Which LLM APIs have a genuinely free tier or free credits in 2026, and what tokens cost once you exceed it. OpenAI, Anthropic, Google Gemini, Mistral, Groq, DeepSeek, Cerebras, OpenRouter, Cohere and xAI compared — free tier limits, rate limits, context windows and per-token pricing.
Published 2026-04-13 · Reviewed 2026-10-15 · 19 providers compared · Figures compiled 2026-04-13, last checked 2026-10-15 · 96 of 97 figures in the tables below come from our records for 1,493 developer tools · 62 pricing changes tracked
LLM API pricing: 19 providers across four categories — frontier labs, inference providers, open-source hosts, and specialized services. OpenAI and Anthropic price their top model identically: GPT-6 Astra and Claude Fable 5.1 are both $10/$50 per M tokens. Gemini 3.8 Flash is $0.75/$3.75 through December 31, 2026 and $1.50/$7.50 after it. Mistral Medium 3.5 is $1.50/$7.50. Google's Gemini free tier covers the 3.x Flash and Flash-Lite models; 3.1 Pro Preview is paid-only. DeepSeek's deepseek-flash offers 1M context at $0.30/M input at peak and $0.15/M off-peak. Groq offers genuinely free inference at thousands of tokens/second.
Key trends: Inference providers (Groq, OpenRouter) are commoditizing open-source model access — free tiers with no credit card required. xAI retired Grok 4.1 Fast on 2026-05-15; requests to its model names now go to grok-4.3 at $1.25/M input and $2.50/M output (under 200k prompt tokens). The gap between frontier and open-source quality is narrowing, making the price delta harder to justify for many use cases.
This guide covers: pricing tables, provider breakdowns, free tier analysis, pricing gotchas, recent changes, and options by use case.
Most generous: Groq — 30 RPM free · Mistral AI — Free plan includes monthly API usage whose amount the pricing page has not stated since 2026-10-07 · Best for prototyping: OpenRouter — 25+ free models · Cloudflare Workers AI — 10K neurons/day · Free for light use: LLM7.io — 100,000 tokens/24h with a free token
Hover rows to highlight. Click provider names for full vendor profiles.
| Provider | Recorded Tier | Paid Rate |
|---|---|---|
| OpenAI | Pay-as-you-go | $0.10/$0.50 (gpt-6-luna) – $10.00/$50.00 (gpt-6-astra) per MTok |
| Anthropic | Pay-as-you-go | $0.10/$0.50 (Claude Haiku 5.5) – $10/$50 (Claude Fable 5.1) per MTok |
| Google Gemini | Free (Reduced) | $0.25/$1.50 (Gemini 3.1 Flash-Lite) – $2/$12 (Gemini 3.1 Pro Preview) per MTok |
| Mistral AI | Free | $0.1/$0.1 (Ministral 3) – $1.5/$7.5 (Mistral Medium 3.5) per MTok |
| Cohere | Free | $0.0375/$0.15 (Command R7B), $0.15/$0.60 (Command R) per MTok |
| xAI (Grok) | Pay-as-you-go | $1.00/$2.00 (grok-build-0.1), $1.25/$2.50 (grok-4.3), $2.00/$6.00 (grok-4.7) per MTok |
| Groq | Free | $0.075/$0.30 (gpt-oss-20b), $0.15/$0.60 (gpt-oss-120b), $0.80/$4.00 (Qwen3.8 27B) per MTok |
| Cerebras | Trial | $0.35/$0.75 (GPT OSS 120B), $0.99/$1.49 (Qwen 3.8 27B) per MTok |
| OpenRouter | Free | — vendor pricing |
| NVIDIA NIM | Free | — vendor pricing |
| SiliconFlow | Free | $0.04/$0.18 (gpt-oss-20b), $0.05/$0.45 (gpt-oss-120b), $0.15/$0.60 (DeepSeek-V4.1-Flash) per MTok |
| Hugging Face | Free | — vendor pricing |
| Replicate | Trial | $3.75/$10 (DeepSeek-R1) per MTok |
| Baseten | Basic (Free Credits) | $0.15/$0.50 (GLM-5.3-Flash), $1.40/$4.40 (GLM-5.3) per MTok |
| Cloudflare Workers AI | Free | — vendor pricing |
| DeepSeek | Pay-as-you-go | $0.30/$1.20 (DeepSeek-V4.1-Flash), $1.32/$3.96 (DeepSeek-V4-Pro) per MTok |
| GitHub Models | Retired | — |
| LLM7.io | Free | — vendor pricing |
| Ollama Cloud | Free | — vendor pricing |
LLM API providers fall into four categories. Each optimized for different use cases, budgets, and requirements.
A date in a heading below cites a record we hold, not a property of the vendor it names. These figures were compiled on 2026-04-13; a dated badge marks a card whose vendor we have recorded a change for since, and the date moves to the newer record on the day that record takes effect. Every record we hold for any vendor named below is in the full change log.
The AI labs that train and serve their own flagship models. Higher prices, cutting-edge capabilities, proprietary architectures.
AI API platform. One model is priced Free in OpenAI's own table: the moderation model omni-moderation-latest. No GPT model is priced free. GPT prices per 1M tokens (Standard, short context): gpt-6-astra $10.00/$50.00; gpt-6.1-sol $2.00/$10.00; gpt-6-sol $2.00/$10.00; gpt-6-luna $0.10/$0.50; gpt-5.6-sol $4.00/$20.00 (a promotional price, available at least through November 21, 2026); gpt-5.6-terra $2.00/$12.00; gpt-5.6-luna $0.20/$1.20. Batch and Flex processing halve these prices. Embeddings from $0.02 per 1M tokens. Three further free amounts are sub-quotas inside paid tools: 1 GB of File search storage (then $0.10/GB per day), 1 GB per account per month of ChatKit upload storage, and search content tokens from the web search preview tool on non-reasoning models.
Claude API with usage-based pricing per million tokens (input/output). Claude Fable 5.1 $10/$50. Claude Opus 5.5 $4/$20. Claude Sonnet 5.5 $2/$10. Claude Sonnet 5 $2/$10. Claude Haiku 5.5 $0.10/$0.50 (for prompts up to 100,000 tokens; longer prompts cost five times as much). Claude Haiku 4.5 $1/$5. The Batch API gives a 50% discount on input and output tokens. New users receive a small amount of free credits to test the API.
Free tier: Gemini 3.8 Flash, 3.6 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite, Gemini 3 Flash Preview, Gemini Embedding 2 and Gemma 4 are free of charge. Also free of charge are Gemini 3.8 Live (real-time voice), Gemini 3.5 Live Translate (translation), Gemini 3.5 Transcribe (speech-to-text), Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS (text-to-speech). Gemini 4 Argon (gemini-4-argon), Gemini 3.1 Pro Preview and Gemini Omni Flash (gemini-omni-1.1-flash) are paid only. Gemini 4 Argon, generally available since 2026-10-13, is not available on the Gemini API free tier. According to ai.google.dev/gemini-api/docs/google-ai-plans, "Gemini 4 Argon is not available on the Free plan." Gemini 4 Argon is available through the Gemini API on the paid tier. "Direct use of the Gemini API (such as using API keys or external applications) is billed and managed separately," per ai.google.dev/gemini-api/docs/google-ai-plans. Google AI Pro and Ultra include Gemini 4 Argon in Google AI Studio and Antigravity (Pro: non-trial subscriptions only), per ai.google.dev/gemini-api/docs/google-ai-plans and antigravity.google/docs/models. No Gemini image model has a free tier in the API. Nano Banana 2.1 (gemini-nano-banana-2.1) became generally available on 2026-10-06. It costs $0.0336 per 1K image, $0.0504 per 2K image and $0.113 per 4K image. The Batch API halves those prices. Nano Banana Pro costs $0.134 per 1K or 2K image and $0.24 per 4K image. Google publishes no free-tier rate limits; each project's limits are shown in Google AI Studio. Developers report 429 errors showing 20 requests per day for Gemini 3.x Flash models and 500 requests per day for 3.x Flash-Lite models as of September 2026. Rate limits apply per project, not per API key. Daily request quotas reset at midnight Pacific time, per Google's rate-limits page. Since 2026-09-18 Google serves the Gemini 2.5 models only to users who have used them before, and points new projects to 3.5 Flash-Lite or 3.8 Flash. If your app calls gemini-3.7-flash, Google deprecated it on 2026-10-08. Requests to gemini-3.7-flash are routed automatically to gemini-3.8-flash. Google says to update the model string to gemini-3.8-flash. If your app calls Gemini 3.5 Flash (gemini-3.5-flash), Google deprecated it on 2026-10-08. Requests to gemini-3.5-flash are routed automatically to gemini-3.6-flash. Google says to update the model string to gemini-3.6-flash or gemini-3.8-flash. If your app calls gemini-omni-flash-preview, it can be called until 2026-10-22, when Google shuts it down. Google's replacement is gemini-omni-1.1-flash. Gemini 3.1 Flash-Lite (gemini-3.1-flash-lite) can be called until 2027-05-07, when Google shuts it down. Google's replacement is gemini-3.5-flash-lite. Paid, per million tokens (input/output; output prices include thinking tokens): Gemini 4 Argon $2/$10 through 2026-11-30 (context caching $0.10, then $0.20, implicit caching only); Gemini 3.8 Flash $0.75/$3.75 until 2026-12-31, then $1.50 input and $7.50 output from 2027-01-01 (3.6 Flash: the same prices and dates); Gemini 3.5 Flash-Lite $0.30/$2.50; Gemini 3.1 Flash-Lite $0.25/$1.50; Gemini 3.1 Pro Preview $2/$12 (prompts up to 200K tokens; prompts over 200K tokens cost $4 input and $18 output). From December 1, 2026, Gemini 4 Argon costs $4 per million input tokens and $20 per million output tokens for Standard, $2 input and $10 output for Batch API, $2 input and $10 output for Flex, and $7.20 input and $36 output for Priority. The Batch API costs half the standard price; Gemini 3.8 Flash is $0.375 input and $1.875 output per million tokens through 2026-12-31. Gemini 4 Argon on the Batch API is $1 input and $5 output through 2026-11-30. Google also sells two tiers, Flex and Priority, both in Preview. A request picks one with the service_tier field ("flex" or "priority"); without it the request is Standard. Gemini 4 Argon on Flex is $1 input and $5 output through 2026-11-30; on Priority it is $3.60 input and $18 output through 2026-11-30. "The Gemini Flex API is an inference tier that offers a 50% cost reduction compared to standard rates, in exchange for variable latency and best-effort availability" (ai.google.dev/gemini-api/docs/flex-inference). "Priority inference is priced at 75-100% more than the standard API and billed per token" (ai.google.dev/gemini-api/docs/priority-inference). The Gemini API Contributor Tier is an invite-only private preview. Eligible users will be invited by email. It offers half the Standard price for Gemini 4 Argon, Gemini 3.1 Pro Preview, Gemini 3.8 Flash, and Gemini 3.5 Flash-Lite. Google uses the data from those models to improve its products. "Please don't submit sensitive or personal information" (ai.google.dev/gemini-api/docs/logs-policy). Grounding with Google Search is supported on the free tier for Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking and Gemini 3.1 Flash Live Preview. It is not available on the free tier for Gemini 4 Argon or the Gemini 3.x Flash, Flash-Lite and Pro models. On the paid tier, 5,000 search requests a month are free, shared across Gemini 3 and newer models, then $14 per 1,000. Accounts opened after 2026-03-02 cannot spend the $300 Google Cloud welcome credit on the Gemini API.
Mistral's Free plan includes monthly usage shared across Studio, the API and Vibe Code. The pricing page no longer states the amount as of 2026-10-07; it previously listed $10 a month in API credits. When it is used up, usage can stop until the next month unless pay-as-you-go is turned on. Mistral Pro costs $14.99 a month and includes $25.50 a month in API credits, shared across the API, Studio and Vibe Code. It included $30 a month until 2026-10-07. Verified students who have not used Vibe or Le Chat before can take the Education plan for $5.99 a month, for up to 12 months, with $15 a month in API credits. API keys work with no credit card, within rate limits. The plan also has limited messages, web searches and coding sessions in Vibe. API prices: Ministral 3 (3B) $0.1/$0.1 (per 1M tokens); Mistral Small 4 $0.15/$0.6 (per 1M tokens); Mistral Large 3 $0.5/$1.5 (per 1M tokens); Mistral Large 4 (mistral-large-4; public preview since 2026-10-06, 1M-token context) $1.36/$4.18 (per 1M tokens); its launch discount to half that ended on 2026-10-20; Mistral Medium 3.5 $1.5/$7.5 (per 1M tokens); Codestral $0.3/$0.9 (per 1M tokens). To call Mistral Large 4, use the model name mistral-large-4; mistral-large-latest still points to Mistral Large 3 (mistral-large-2512). Mistral Embed is $0.1 per 1M input tokens, and Codestral Embed is $0.15 per 1M input tokens. OCR 4.1 costs $4 per 1,000 pages. Mistral Moderation 2 is free. Speech: Voxtral 3 models, released 2026-10-20: Voxtral Mini Transcribe 3 (batch) costs $0.0035 per minute of audio, Voxtral Small Transcribe 3 $0.004 per minute, Voxtral Mini Transcribe Realtime 3 $0.008 per minute, and Voxtral Speech 3 (text-to-speech) $13 per 1M characters. The earlier models are still offered: Voxtral Mini Transcribe 2 $0.003 per minute, Voxtral Mini Transcribe Realtime $0.006 per minute, and Voxtral TTS (text-to-speech) $16 per 1M characters. Mistral's voice agent API costs $0.04 per minute with Mistral Small 4 for reasoning, and $0.08 per minute with any other reasoning model. Mistral retired these models in 2026, and a request to a retired model fails with a 404 error: Mistral Large 2.1 (mistral-large-2411), Pixtral Large (pixtral-large-2411) and Devstral Medium 1.0 (devstral-medium-2507) on 2026-05-31, replaced by Mistral Medium 3.5; Devstral Small 1.1 (devstral-small-2507) on 2026-05-31, replaced by Mistral Small 4; Voxtral Mini and Voxtral Mini Transcribe (voxtral-mini-2507) on 2026-05-31, replaced by Voxtral Mini Transcribe 2; OCR 2 (mistral-ocr-2505) on 2026-05-31 and OCR 4.0 (mistral-ocr-4-0) on 2026-09-30, replaced by OCR 4.1; Mistral Moderation (mistral-moderation-2411) on 2026-06-30, replaced by Mistral Moderation 2; Mistral Small 3.2 (mistral-small-2506) and Magistral Small 1.2 (magistral-small-2509) on 2026-07-31, replaced by Mistral Small 4; Mistral Nemo 12B (open-mistral-nemo-2407) on 2026-07-31, replaced by Ministral 3 8B; Devstral 2 (devstral-2512) and Magistral Medium 1.2 (magistral-medium-2509) on 2026-07-31, and Mistral Medium 3 and 3.1 (mistral-medium-2505, mistral-medium-2508) on 2026-08-31, replaced by Mistral Medium 3.5. Mistral's table at docs.mistral.ai/models also lists the Labs models it retired. Z.ai GLM 5.2 (zai-glm-5-2) retires on 2026-10-31, replaced by GLM 5.3. Batch processing is half price, and cached input tokens cost up to 90% less. Mistral publishes no rate-limit figures. Each organization's per-model limits (tokens per minute, requests per second) appear in Admin Panel > API > Limits. The tools of Mistral's MCP server are grouped by Studio capability: Skills, Workflows, Connectors, Observability, Evaluations, Organization and Apps.
AI model API for Command chat models, Embed, Rerank, Transcribe and Parse. Every account starts with a Trial API key: calls made with it are free and limited to 1,000 API calls a month. Trial rate limits: 20 requests a minute per Chat model, 2,000 Embed inputs a minute and 10 Rerank requests a minute. Trial keys can use all of Cohere's models and APIs. Production keys are pay-as-you-go. Prices per 1M tokens: Command R7B $0.0375/$0.15; Command R $0.15/$0.60. The pricing page lists Command A+ (Apache 2.0) at $0 through an API key and as a model download.
Grok API, pay as you go: sign up at console.x.ai, then load it with credits. grok-4.7 $2.00/$6.00 (per 1M tokens, under 200k prompt tokens). grok-4.3 $1.25/$2.50 (under 200k prompt tokens). grok-build-0.1 $1.00/$2.00 (under 200k prompt tokens). grok-4.6 and grok-4.5 cost the same as grok-4.7 ($2.00 input and $6.00 output per 1M tokens under 200k prompt tokens), and grok-4.20 (0309 reasoning, non-reasoning and multi-agent) the same as grok-4.3 ($1.25 input and $2.50 output). A request whose prompt reaches 200k tokens is billed at the long-context rate for all its tokens, twice the standard rate: grok-4.7 and grok-4.6 go from $2.00 input and $6.00 output per 1M tokens to $4.00 and $12.00; grok-4.3 and grok-4.20 from $1.25 and $2.50 to $2.50 and $5.00; grok-build-0.1 from $1.00 and $2.00 to $2.00 and $4.00. Cached input costs $0.50 on grok-4.7 and grok-4.6, $0.30 on grok-4.5, and $0.20 on grok-4.3, grok-4.20 and grok-build-0.1. These cached rates also double from 200k prompt tokens. Context windows: 500k tokens for grok-4.7, grok-4.6 and grok-4.5; 1M for grok-4.3 and grok-4.20; 256k for grok-build-0.1. The US endpoint costs 10% more than the global endpoint. Priority Processing (service_tier "priority" on Chat Completions and the Responses API) bills text requests at 2x standard rates. A request xAI's system finds in violation of its usage guidelines is still charged for its generation; one caught before generation in the Responses API costs a $0.05 violation fee. On 2026-05-15 xAI retired eight model names, which still resolve, so code that calls them keeps working: requests to Grok 4.1 Fast (grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning), Grok 4 Fast (grok-4-fast-reasoning, grok-4-fast-non-reasoning), grok-4-0709 and grok-3 are served by grok-4.3 (low reasoning effort for the reasoning names and grok-4-0709, none for the non-reasoning names and grok-3) and billed at grok-4.3 rates; grok-code-fast-1 is served by grok-build-0.1; grok-imagine-image-pro is served by grok-imagine-image-quality. grok-voice-transcribe-1.0 reached end of life on 2026-10-02; its requests are served by grok-voice-transcribe-2.0 at the same price. Image generation: grok-imagine-image $0.02 per image; grok-imagine-image-2.0 from $0.04 per 1K image at low quality. grok-imagine-image-quality ($0.05 per 1K image) is retired on 2026-11-02, when its requests are served by grok-imagine-image-2.0 at low quality. Voice API: speech-to-text costs $0.10 per hour of audio (REST) or $0.20 per hour (streaming), text-to-speech $15.00 per 1M characters, and speech-to-speech with grok-voice-3.0 $0.08 a minute ($4.80 an hour). grok-voice-3.0 replaced grok-voice-2.0 on 2026-10-14 at the same price; grok-voice-2.0 is deprecated, and the grok-voice-latest alias now points to grok-voice-3.0. Video generation: grok-imagine-video-1.5 $0.08 per second at 480p, $0.14 at 720p, $0.25 at 1080p. grok-imagine-video-1.5-lite $0.02 per second at 480p, $0.03 at 720p, $0.14 at 1080p. grok-imagine-video $0.05 per second at 480p, $0.07 at 720p. Batch discounts: grok-4.3, grok-4.20-0309-reasoning, grok-4.20-0309-non-reasoning, and grok-4.20-multi-agent-0309 get 20% off. Other models have no batch discount. Server-side tools cost extra, per call, on top of tokens: web_search, code_execution and attachment_search $5 per 1,000 calls, collections_search $2.50 per 1,000, and x_search $5 per 1,000 posts and $10 per 1,000 profiles it returns. view_image, view_x_video and remote MCP tools are billed for their tokens only.
Platforms that serve open-source models on optimized hardware. Cheaper than frontier labs, focused on speed and cost efficiency.
Fast LLM inference on Groq's LPU hardware. Free plan: gpt-oss-120b, gpt-oss-20b and Qwen3.8 27B, each at 30 requests a minute, 1,000 requests and 200,000 tokens a day, plus Whisper speech-to-text at 2,000 requests a day. Each of the three free models is also limited to 8,000 tokens a minute, and Groq applies whichever limit is reached first. Whisper is also limited to 20 requests a minute, 7,200 audio seconds an hour and 28,800 audio seconds (8 hours of audio) a day. On the free plan, gpt-oss-safeguard-20b is limited to 3 requests a minute, 1,000 requests a day, 2,000 tokens a minute and 200,000 tokens a day. Qwen3.8 27B, gpt-oss-safeguard-20b, Orpheus Arabic Saudi and Orpheus V1 English are Preview models. Groq says Preview models "are intended for evaluation purposes only and should not be used in production environments as they may be discontinued at short notice". Llama 3.3 70B and Llama 3.1 8B left the free and developer plans on 2026-08-16 and are Enterprise-only. Orpheus text-to-speech (English and Saudi Arabic) is also on the free plan, at 10 requests a minute, 100 requests and 3,600 tokens a day. Developer plan prices per 1M tokens: gpt-oss-120b $0.15/$0.60; gpt-oss-20b $0.075/$0.30; Qwen3.8 27B $0.80/$4.00. The Batch API costs 50% less and returns results within 24 hours to 7 days, without impacting standard rate limits. When requests to gpt-oss-20b, gpt-oss-120b or gpt-oss-safeguard-20b start with the same text as a recent request, Groq caches that part automatically and bills those input tokens 50% less. The batch discount does not stack with prompt caching discounts; all batch tokens are billed at the batch rate regardless of cache status. Orpheus text-to-speech costs $22 per 1M characters in English and $40 in Saudi Arabic. Groq's API is OpenAI-compatible at the base URL https://api.groq.com/openai/v1. The free models' API model IDs are openai/gpt-oss-120b, openai/gpt-oss-20b and qwen/qwen3.8-27b.
Ultra-fast LLM inference API — no permanent free tier. New accounts get $5 in free credits after adding a verified payment method, expiring 30 days after they are granted. Free Trial limits on gpt-oss-120b and qwen-3.8-27b: 5 requests/min and 1M tokens/day. Pay as you go after: GPT OSS 120B $0.35/$0.75 per M tokens, Qwen 3.8 27B $0.99/$1.49 per M tokens. Cerebras's public API now serves only gpt-oss-120b and qwen-3.8-27b, and a request to a model ID it has deprecated fails with 404 model not found. Model IDs Cerebras deprecated in 2026: zai-glm-4.6 on 2026-01-20; llama-3.3-70b on 2026-02-16, replaced by gpt-oss-120b; qwen-3-32b on 2026-02-16, replaced by gpt-oss-120b; llama3.1-8b on 2026-05-27, replaced by gpt-oss-120b; qwen-3-235b-a22b-instruct-2507 on 2026-05-27, replaced by gpt-oss-120b; zai-glm-4.7 on 2026-08-17, with no replacement named; gemma-4-31b on 2026-09-03, replaced by qwen-3.8-27b (Gemma 4 31B stays available on Dedicated Inference). Multi-thousand tokens/sec inference speed.
AI model router. Free plan: 50 requests a day, no BYOK. Free models are capped at 20 requests a minute; accounts that have bought at least $10 of credits get 1,000 free-model requests a day. OpenRouter's pricing page says "25+ free models" and "4 free providers"; its model list (openrouter.ai/api/v1/models) had 17 model IDs ending in ":free" on 2026-10-20, and that list changes from week to week. The Standard plan (pay-as-you-go) charges a 5.5% fee on credit purchases and includes BYOK up to $25,000 of list-price inference a month with no fees, 5% after. On 2026-08-19 OpenRouter announced it is joining Stripe. OpenRouter says it will continue to operate as it is, with the same name, product and roadmap, and that existing integrations do not change.
Free inference endpoints on build.nvidia.com: models marked Free Endpoint, including DeepSeek V4.1 Flash, GLM 5.3 Flash, gpt-oss-20b, Gemma 4 31B and NVIDIA Nemotron 3 Ultra, can be called at no cost. Their model IDs are deepseek-ai/deepseek-v4.1-flash, z-ai/glm-5.3-flash, openai/gpt-oss-20b, google/gemma-4-31b-it and nvidia/nemotron-3-ultra-550b-a55b. Up to 40 requests per minute; limits may vary by model, and traffic from other users may cause throttling. NVIDIA's API Trial Terms allow free use for testing and evaluation only, not production. NVIDIA posts an end-of-life date on each hosted model's page on build.nvidia.com; after that date, requests for the model return HTTP 410 Gone. GLM 5.3 (z-ai/glm-5.3) can be called until 2026-10-27; NVIDIA's page says it is no longer supported after 2026-10-28. This date does not apply to GLM 5.3 Flash (z-ai/glm-5.3-flash). NVIDIA stopped supporting NVIDIA Nemotron 3 Super (nvidia/nemotron-3-super-120b-a12b) after 2026-10-19, and its catalogue no longer lists the model (checked 2026-10-21). Nemotron 3.5 Super VL (nvidia/nemotron-3.5-super-vl-120b-a12b), an early-access model, is a Free Endpoint. NVIDIA retired Kimi K3 (moonshotai/kimi-k3) on 2026-10-15. Other model IDs NVIDIA retired in 2026 include these. DeepSeek: deepseek-ai/deepseek-v4-pro and deepseek-ai/deepseek-v4-flash on 2026-08-07, deepseek-ai/deepseek-v4-pro-0813 on 2026-09-14 and deepseek-ai/deepseek-v4-flash-0731 on 2026-09-21. Kimi: moonshotai/kimi-k2-instruct and moonshotai/kimi-k2-thinking on 2026-05-12. GLM: z-ai/glm4.7 on 2026-05-14, z-ai/glm-5.1 on 2026-07-02 and z-ai/glm-5.2 on 2026-08-21. Qwen: qwen/qwen2.5-coder-32b-instruct on 2026-05-12, qwen/qwen3-next-80b-a3b-thinking on 2026-05-21, qwen/qwen3-coder-480b-a35b-instruct on 2026-06-11, qwen/qwen3.5-122b-a10b on 2026-07-20, and qwen/qwen3.5-397b-a17b and qwen/qwen3-next-80b-a3b-instruct on 2026-07-27. MiniMax: minimaxai/minimax-m2.5 on 2026-05-12, minimaxai/minimax-m2.7 on 2026-07-27 and minimaxai/minimax-m3 on 2026-09-09. gpt-oss: openai/gpt-oss-120b on 2026-09-03. Llama: meta/llama-4-maverick-17b-128e-instruct on 2026-07-27, and meta/llama-3.3-70b-instruct, meta/llama-3.1-8b-instruct, meta/llama-3.1-70b-instruct, meta/llama-3.2-1b-instruct and meta/llama-3.2-3b-instruct on 2026-08-26. Nemotron: nvidia/llama-3.3-nemotron-super-49b-v1, nvidia/llama-3.3-nemotron-super-49b-v1.5, nvidia/llama-3.1-nemotron-nano-8b-v1, nvidia/llama-3.1-nemotron-nano-vl-8b-v1, nvidia/nvidia-nemotron-nano-9b-v2, nvidia/nemotron-nano-12b-v2-vl and nvidia/nemotron-mini-4b-instruct on 2026-08-26, and nvidia/nemotron-3-nano-30b-a3b on 2026-09-01. Mistral: mistralai/mistral-large-3-675b-instruct-2512 on 2026-07-23, mistralai/mixtral-8x7b-instruct-v0.1 on 2026-07-27 and mistralai/mistral-nemotron on 2026-09-28. Gemma: google/gemma-3-27b-it on 2026-05-12, and google/gemma-3n-e4b-it and google/gemma-3n-e2b-it on 2026-07-27. Phi: microsoft/phi-4-mini-instruct and microsoft/phi-4-multimodal-instruct on 2026-07-15. Others: stepfun-ai/step-3.5-flash, bytedance/seed-oss-36b-instruct, sarvamai/sarvam-m and abacusai/dracarys-llama-3.1-70b-instruct on 2026-07-27. Embedding models: nvidia/llama-3.2-nv-embedqa-1b-v2 and nvidia/llama-3.2-nemoretriever-300m-embed-v1 on 2026-05-18, and nvidia/nv-embedqa-e5-v5, nvidia/nv-embed-v1, nvidia/llama-nemotron-embed-1b-v2 and baai/bge-m3 on 2026-08-25; nvidia/nemotron-3-embed-1b is a Free Endpoint embedding model.
Pay-as-you-go model inference. siliconflow.cn, run by a Beijing company, lists these models at no charge: the chat models Qwen3-8B, Qwen2.5-7B-Instruct (Free), Qwen3.5-4B, GLM-4-9B-0414, GLM-Z1-9B-0414, DeepSeek-R1-0528-Qwen3-8B (Free) and Xing4.0-29B; the translation model Hunyuan-MT-7B; the OCR models PaddleOCR-VL-1.5 and DeepSeek-OCR; the embedding models bge-m3, bge-large-zh-v1.5 and bge-large-en-v1.5; the reranker bge-reranker-v2-m3; Kolors image generation; and the speech recognition models XingChenAGI/XingChenGSR-V1.0, XingChenAGI/XingChenASR-V3.2-Ultra, XingChenAGI/XingChenASR-Diarize-V3.0, XingChenAGI/XingChenASR-V3.2 and FunAudioLLM/SenseVoiceSmall. Qwen3-ASR-1.7B is paid at ¥0.00022 per second of audio. The Pro versions are paid: Qwen2.5-7B-Instruct (Pro), bge-m3 (Pro) and bge-reranker-v2-m3 (Pro). All other models are paid too, including DeepSeek-V4-Flash, DeepSeek-V4-Pro and GLM-5.3. Using all the free models requires real-name verification, and online personal verification accepts only Chinese-issued documents, such as a resident ID card or a Foreign Permanent Resident ID Card. The international site, siliconflow.com, is run by SiliconFlow Labs Pte. Ltd. under Singapore law, is not offered in mainland China, and gives $1 in free credits to start. Its paid rates per million input/output tokens: gpt-oss-20b $0.04/$0.18; gpt-oss-120b $0.05/$0.45; DeepSeek-V4.1-Flash $0.15/$0.60.When we read the page we cite for this offer on 2026-09-16, we found the page did not mention this offer, which is not evidence it ended, and we have read it again since, on 2026-10-01, without confirming the terms above, so we cannot confirm these terms today.
Platforms for deploying and serving any open-source model. Pay for compute, bring your own model.
ML model hub. Since 2026-10-07, free accounts get no included Inference Providers credits: free users buy credits or upgrade to PRO to call Inference Providers, which serves 200+ models. PRO ($9/month) gets $2.00 a month. Free accounts get 100GB of private storage and best-effort public storage; substantial storage needs PRO, Team or Enterprise.
ML model hosting and inference. A new account can run the models in Replicate's Try for Free collection without buying credit, for a limited number of runs that Replicate does not state; after that you add billing and buy credit. Replicate says these models are not free forever. Most models bill by the second for the hardware they run on, such as $0.000025/sec on CPU Small and $0.001525/sec on one H100. Official models bill per output instead, such as FLUX 1.1 Pro at $0.04 per image; per million input/output tokens: DeepSeek-R1 $3.75/$10.
ML model deployment platform. New workspaces receive credits for testing and deployment; Baseten does not state the amount. Basic plan: $0 per month, pay as you go. Dedicated deployments bill per minute. Model APIs bill per 1M tokens: GLM-5.3 $1.40/$4.40. GLM-5.3-Flash $0.15/$0.50.When we read the page we cite for this offer on 2026-09-10, we found a change we could not reconcile with the terms we publish, and we have read it again since, on 2026-10-04, without confirming them, so we cannot confirm these terms today.
Workers AI runs AI models on Cloudflare's global network. Every account gets 10,000 Neurons per day at no charge, reset daily at 00:00 UTC; on the Workers Free plan, requests beyond that fail until the reset. On the Workers Paid plan (from $5 per month), usage above 10,000 Neurons per day costs $0.011 per 1,000 Neurons. Seven models, including @cf/moonshotai/kimi-k2.6, @cf/zai-org/glm-5.3 and @cf/deepseek-ai/deepseek-v4-pro-0813, require the Workers Paid plan or prepaid AI Gateway credits. On 2026-05-30, Cloudflare deprecated 18 models and told users to remove references to them: @cf/meta/llama-3.1-8b-instruct, @cf/meta/llama-3.1-8b-instruct-awq, @cf/meta/llama-3.1-70b-instruct, @cf/meta/llama-3-8b-instruct, @cf/meta/llama-3-8b-instruct-awq, @hf/meta-llama/meta-llama-3-8b-instruct, @cf/meta/llama-2-7b-chat-int8, @cf/meta/llama-2-7b-chat-fp16, @cf/mistral/mistral-7b-instruct-v0.1, @hf/mistral/mistral-7b-instruct-v0.2, @hf/google/gemma-7b-it, @cf/google/gemma-3-12b-it, @hf/nousresearch/hermes-2-pro-mistral-7b, @cf/microsoft/phi-2, @cf/defog/sqlcoder-7b-2, @cf/unum/uform-gen2-qwen-500m, @cf/facebook/bart-large-cnn and @cf/moonshotai/kimi-k2.5. On 2026-10-15, Cloudflare's model catalog still lists @cf/meta/llama-3.1-8b-instruct-fp8, @cf/meta/llama-3.3-70b-instruct-fp8-fast and the -lora variants, but no longer lists @cf/meta/llama-3.1-8b-instruct-fast. Cloudflare's suggested replacements are @cf/zai-org/glm-4.7-flash, @cf/google/gemma-4-26b-a4b-it and @cf/moonshotai/kimi-k2.6. Since 2026-05-30, requests to @cf/moonshotai/kimi-k2.5 go to @cf/moonshotai/kimi-k2.6, which costs more and requires the Workers Paid plan or prepaid AI Gateway credits; without them, requests get HTTP 403 (error 5035).
Providers with unique positioning — free tiers, self-hosting, or niche model access.
Pay-as-you-go LLM API with no free tier, billed from a prepaid balance. DeepSeek-V4.1-Flash (deepseek-flash): $0.30/M input, $1.20/M output at peak. DeepSeek-V4-Pro (deepseek-v4-pro): $1.32/M input, $3.96/M output at peak. Cache-hit input is $0.006/M on Flash and $0.044/M on Pro. Off-peak rates are half; peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays. 1M context, up to 384K output. China-based. To call the API today, use deepseek-flash or deepseek-v4-pro: DeepSeek retired the older model names deepseek-chat and deepseek-reasoner on 2026-07-24 (its change log of 2026-04-24 says they "will be discontinued in three months (2026-07-24)"). Until then both ran on the Flash model, deepseek-chat in non-thinking mode and deepseek-reasoner in thinking mode, so deepseek-flash replaces either.
Free tier: none. GitHub Models's offer is recorded as Retired.
LLM inference gateway. With a free token from dash.llm7.io: 1 request a second, 60 a minute, 250 an hour and 100,000 tokens per 24 hours, counting input and output tokens; LLM7 says free quotas may be reduced without notice. Free access covers turbo-tier models not marked usage-only, such as GLM-5.3-Flash and codestral-latest. Other models, including all image and speech-to-text models, need Pro ($12/mo) or a paid balance. LLM7's limits page no longer lists limits for requests without a token. A text request without a token still worked on 2026-10-19.
Cloud-hosted Ollama for running open-source LLMs — The Free plan costs $0 and includes starter usage credits. Buying usage credits unlocks all models. It includes 1 concurrent request. Pro costs $20 a month with $60 of usage credits, Max $100 a month with $300. Team, in early access, costs $500 a month for unlimited users and includes $1,000 of usage credits a month, shared across the team. Per million input and output tokens, glm-5.3 costs $1.40 and $4.40, deepseek-v4-pro $1.32 and $3.96, and kimi-k3 $3.00 and $15.00. Usage is billed per million tokens at each model's rate, for example gpt-oss:120b at $0.15 input and $0.60 output, and some models, such as deepseek-v4.1-flash, cost half off-peak (outside 12:00 to 18:00 UTC on weekdays, and all weekend). To use cloud models without installing Ollama, send requests to https://ollama.com/api/chat with an API key from ollama.com in an "Authorization: Bearer" header, using the same model names, such as gpt-oss:120b; OpenAI and Anthropic clients also work.
Free tiers range from genuinely production-viable (Groq) to token giveaways that run out in hours. Here's the honest breakdown.
| Provider | Recorded Tier | Free Tier |
|---|---|---|
| OpenAI | Pay-as-you-go | AI API platform. One model is priced Free in OpenAI's own table: the moderation model omni-moderation-latest. No GPT model is priced free. Full profile |
| Anthropic | Pay-as-you-go | Claude API with usage-based pricing per million tokens (input/output). Claude Fable 5.1 $10/$50. Claude Opus 5.5 $4/$20. Claude Sonnet 5.5 $2/$10. Claude Sonnet 5 $2/$10. Full profile |
| Google Gemini | Free (Reduced) | Free tier: Gemini 3.8 Flash, 3.6 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite, Gemini 3 Flash Preview, Gemini Embedding 2 and Gemma 4 are free of charge. Full profile |
| Mistral AI | Free | Mistral's Free plan includes monthly usage shared across Studio, the API and Vibe Code. The pricing page no longer states the amount as of 2026-10-07; it previously listed $10 a month in API credits. Full profile |
| Cohere | Free | AI model API for Command chat models, Embed, Rerank, Transcribe and Parse. Every account starts with a Trial API key: calls made with it are free and limited to 1,000 API calls a month. Full profile |
| xAI (Grok) | Pay-as-you-go | Grok API, pay as you go: sign up at console.x.ai, then load it with credits. grok-4.7 $2.00/$6.00 (per 1M tokens, under 200k prompt tokens). grok-4.3 $1.25/$2.50 (under 200k prompt tokens). Full profile |
| Groq | Free | Fast LLM inference on Groq's LPU hardware. Free plan: gpt-oss-120b, gpt-oss-20b and Qwen3.8 27B, each at 30 requests a minute, 1,000 requests and 200,000 tokens a day, plus Whisper speech-to-text at 2,000 requests a day. Full profile |
| Cerebras | Trial | Ultra-fast LLM inference API — no permanent free tier. New accounts get $5 in free credits after adding a verified payment method, expiring 30 days after they are granted. Full profile |
| OpenRouter | Free | AI model router. Free plan: 50 requests a day, no BYOK. Free models are capped at 20 requests a minute; accounts that have bought at least $10 of credits get 1,000 free-model requests a day. Full profile |
| NVIDIA NIM | Free | Free inference endpoints on build.nvidia.com: models marked Free Endpoint, including DeepSeek V4.1 Flash, GLM 5.3 Flash, gpt-oss-20b, Gemma 4 31B and NVIDIA Nemotron 3 Ultra, can be called at no cost. Full profile |
| SiliconFlow | Free | Pay-as-you-go model inference. Full profileWhen we read the page we cite for this offer on 2026-09-16, we found the page did not mention this offer, which is not evidence it ended, and we have read it again since, on 2026-10-01, without confirming the terms above, so we cannot confirm these terms today. |
| Hugging Face | Free | ML model hub. Since 2026-10-07, free accounts get no included Inference Providers credits: free users buy credits or upgrade to PRO to call Inference Providers, which serves 200+ models. PRO ($9/month) gets $2.00 a month. Full profile |
| Replicate | Trial | ML model hosting and inference. A new account can run the models in Replicate's Try for Free collection without buying credit, for a limited number of runs that Replicate does not state; after that you add billing and buy credit. Full profile |
| Baseten | Basic (Free Credits) | ML model deployment platform. New workspaces receive credits for testing and deployment; Baseten does not state the amount. Basic plan: $0 per month, pay as you go. Dedicated deployments bill per minute. Full profileWhen we read the page we cite for this offer on 2026-09-10, we found a change we could not reconcile with the terms we publish, and we have read it again since, on 2026-10-04, without confirming them, so we cannot confirm these terms today. |
| Cloudflare Workers AI | Free | Workers AI runs AI models on Cloudflare's global network. Every account gets 10,000 Neurons per day at no charge, reset daily at 00:00 UTC; on the Workers Free plan, requests beyond that fail until the reset. Full profile |
| DeepSeek | Pay-as-you-go | Pay-as-you-go LLM API with no free tier, billed from a prepaid balance. DeepSeek-V4.1-Flash (deepseek-flash): $0.30/M input, $1.20/M output at peak. DeepSeek-V4-Pro (deepseek-v4-pro): $1.32/M input, $3.96/M output at peak. Full profile |
| GitHub Models | Retired | GitHub retired GitHub Models on 2026-07-30, so there is no free tier. Full profile |
| LLM7.io | Free | LLM inference gateway. With a free token from dash.llm7.io: 1 request a second, 60 a minute, 250 an hour and 100,000 tokens per 24 hours, counting input and output tokens; LLM7 says free quotas may be reduced without notice. Full profile |
| Ollama Cloud | Free | Cloud-hosted Ollama for running open-source LLMs — The Free plan costs $0 and includes starter usage credits. Buying usage credits unlocks all models. It includes 1 concurrent request. Full profile |
Token pricing is rarely the full story. These are the costs and limits that surprise developers.
OpenAI's reasoning models generate internal reasoning tokens that are billed as output tokens but aren't visible in the response; OpenAI states no typical ratio. Monitor usage carefully — your bill reflects total tokens, not just visible output. o4-mini and o3-mini shut down in the OpenAI API on 2026-10-23.
Claude's thinking tokens are billed at output rates ($20/M on Opus 5.5, $50/M on Fable 5.1). A complex reasoning task can produce 10K+ thinking tokens before the actual answer. Budget for 2–5x the output tokens you'd expect from a non-thinking request. Fable 5.1, Opus 5.5, Sonnet 5.5 and Sonnet 5 decide their own thinking budget (adaptive thinking) rather than taking one from the request, so the multiplier is harder to cap than it was on the 4.6 generation.
Google's Gemini free tier covers the 3.x Flash and Flash-Lite models. Gemini 3.1 Pro Preview has no free tier.
A 128K context window doesn't mean the model performs well at 128K tokens. Quality degrades in the middle of long contexts ("lost in the middle" problem). For retrieval-heavy tasks, expect effective context of 30–50% of the advertised window. DeepSeek and Gemini's 1M windows are more susceptible to this.
OpenAI and Anthropic set rate limits by usage tier. OpenAI has three paid tiers since 2026-10-06 and moves an organization up as its total credit purchases reach $5 (Build), $100 (Launch) and $500 (Grow). Anthropic places organizations on a tier based on usage history and account standing, and new organizations may start in an Evaluation tier with lower limits. Groq's free tier limits are per-model, so switching models resets your quota.
LLM API pricing is the most volatile in the developer tools space. Here are the changes we've tracked. See full change timeline for all tracked changes.
| Date | Vendor | Change | Impact |
|---|---|---|---|
| effective Oct 20, 2026 | Google Gemini API | The Gemini API keeps serving gemini-2.5-pro, gemini-2.5-flash and gemini-2.5-flash-lite for existing users with no shutdown date. On Gemini Enterprise Agent Platform (formerly Vertex AI) they can be called until October 20, 2026, when Google retires them. Google names gemini-3.8-flash as a replacement for all three. Source ↗ | MEDIUM |
| effective Oct 19, 2026 | NVIDIA NIM | NVIDIA's free API endpoints keep running. NVIDIA Nemotron 3 Super (nvidia/nemotron-3-super-120b-a12b) can be called on NVIDIA's free endpoints until October 19, 2026; NVIDIA's model page says it will no longer be supported after 10/19/2026. Source ↗ | LOW |
| effective Oct 15, 2026 | NVIDIA NIM | NVIDIA's free API endpoints keep running. NVIDIA retired Kimi K3 (moonshotai/kimi-k3) from its free endpoints on build.nvidia.com on October 15, 2026; its model page says it is no longer supported after 10/15/2026. Source ↗ | LOW |
| effective Oct 14, 2026 | OpenAI Codex | ChatGPT and Codex keep running. Only the GPT-5.5 model retires from ChatGPT, ChatGPT Work and Codex on all plans on 2026-10-14. For Codex with ChatGPT sign-in, OpenAI names GPT-6.1 Sol as the replacement on Plus, Pro, Business, Enterprise and Edu plans (its page named GPT-6 Sol when we read it on 2026-10-06) and GPT-6 Luna on Free and Go plans. The OpenAI API is not affected. Source ↗ | MEDIUM |
| discovered Oct 7, 2026 · effective date unknown | LLM7.io | The free tier limits have been significantly reduced. The free tier now allows 60 requests per minute and 100,000 tokens per 24 hours, whereas previously it allowed 10 requests per minute and 500,000 tokens per 24 hours without a token. The free tier with a token now allows 60 requests per minute and 100,000 tokens per 24 hours, whereas previously it allowed 40 requests per minute and 1,000,000 tokens per 24 hours. Source ↗ | HIGH |
| discovered Oct 7, 2026 · effective date unknown | Mistral AI | Mistral's pricing page stopped listing an API credit on the Free plan. It had listed $10 a month. Mistral's docs still say each plan includes monthly usage but give no amount. Source ↗ | MEDIUM |
| effective Oct 7, 2026 | Hugging Face | Hugging Face stopped giving free accounts included monthly Inference Providers credits; free users now have to buy credits or upgrade to PRO to call Inference Providers. Source ↗ | MEDIUM |
| effective Oct 6, 2026 | Groq | Groq cut the free limits of gpt-oss-safeguard-20b from 30 to 3 requests a minute and from 8,000 to 2,000 tokens a minute; 1,000 requests and 200,000 tokens a day are unchanged. The Internet Archive's capture of console.groq.com/docs/rate-limits on 2026-09-28 lists 30 RPM and 8K TPM, and its capture of 2026-10-06 lists 3 RPM and 2K TPM. Groq announced no change. The other free chat models keep 30 requests and 8,000 tokens a minute. Source ↗ | LOW |
| effective Sep 28, 2026 | OpenAI | gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 shut down in the OpenAI API on September 28, 2026. OpenAI names gpt-5.6-terra as the replacement. Source ↗ | LOW |
| effective Sep 24, 2026 | OpenAI | Sora 2 video generation in the OpenAI API shut down on September 24, 2026: OpenAI removed the Videos API and the sora-2 and sora-2-pro models, and names no replacement. Source ↗ | MEDIUM |
| effective Sep 21, 2026 | Groq | Groq keeps running. Groq decommissioned its Compound and Compound Mini systems, model IDs groq/compound and groq/compound-mini, on September 21, 2026, after announcing it on August 24; requests to those IDs return errors. Groq named no replacement. Source ↗ | MEDIUM |
| effective Sep 18, 2026 | Google Gemini API | Since 2026-09-18 Google serves the Gemini 2.5 models only to users who have used them before, and points new projects to 3.5 Flash-Lite or 3.8 Flash. Source ↗ | MEDIUM |
| effective Sep 18, 2026 | On 2026-09-18 Google stopped renewing every remaining standalone Google Developer Program Premium subscription, Workspace accounts included. Monthly subscribers keep Premium benefits until 2026-10-18 and annual subscribers until 2027-09-18; their membership then returns to the standard tier. Google recommends AI Pro or AI Ultra for consumers, and Gemini Enterprise for enterprises. Source ↗ | MEDIUM | |
| effective Sep 14, 2026 | Groq | Groq replaced Qwen3.6 27B (qwen/qwen3.6-27b) with its successor, Qwen3.8 27B (qwen/qwen3.8-27b), on its free and developer plans on September 14, 2026, and shut qwen/qwen3.6-27b down that day; code that calls it needs the new model ID. Source ↗ | MEDIUM |
| effective Sep 10, 2026 | DeepSeek API | On 2026-09-10, DeepSeek released deepseek-flash (V4.1-Flash) and retired deepseek-v4-flash and deepseek-v4-flash-vision-exp. These retired names still route to V4.1-Flash at Flash prices. Flash prices fell from 04:00 UTC; V4 Pro prices did not change. Source ↗ | MEDIUM |
| discovered Sep 9, 2026 · effective date unknown | GitHub | GitHub Copilot access for students is now limited. Previously offering full access, it now offers a plan with 'limited chat and agent usage with models available through auto model selection only'. Additionally, new Copilot Student signups are paused until 2026-04-20 due to infrastructure strain. Unconfirmed. We could not check this change against an archived copy of GitHub's page, so it does not count toward the rating. Source ↗ | MEDIUM |
| discovered Sep 2, 2026 · effective date unknown | Anthropic API | Anthropic API's own page now redirects to platform.claude.com, and the lineup behind it advanced. Claude Fable 5.1 is $10/$50 per MTok, Opus 5 is $5/$25, Sonnet 5 is $2/$10 and Haiku 4.5 is $1/$5. Anthropic also states that Sonnet 5's $2/$10 introductory pricing is now the standard price and the rise to $3/$15 scheduled for September 1, 2026 will not occur. Source ↗ | HIGH |
| effective between Aug 27, 2026 and Sep 1, 2026 | Ollama Cloud | Ollama's Free plan now includes starter usage credits and starter models only; buying usage credits unlocks the other cloud models. Before, Free included cloud models with a usage limit. Source ↗ | HIGH |
| effective Aug 26, 2026 | OpenAI | Assistants API deprecated, full shutdown August 26, 2026. Developers must migrate to Responses API + Conversations API Source ↗ | HIGH |
| effective Aug 16, 2026 | DeepSeek API | At 16:00 UTC on 2026-08-16, DeepSeek introduced peak and off-peak pricing, with off-peak at half the peak rate. V4 Flash and V4 Pro output prices rose, with Flash from $0.28 to $1.32 at peak and Pro from $0.87 to $3.96 at peak. Source ↗ | HIGH |
These are announced changes that have not taken effect yet, listed soonest first.
| Date | Vendor | Change | Impact |
|---|---|---|---|
| effective Oct 22, 2026 | Google Gemini API | The Gemini API keeps running. Gemini Omni Flash Preview (gemini-omni-flash-preview) and the Veo 3.1 preview models (veo-3.1-generate-preview, veo-3.1-fast-generate-preview and veo-3.1-lite-generate-preview) can be called until October 22, 2026, when Google shuts them down. Google names gemini-omni-1.1-flash as the replacement for all four. Source ↗ | LOW |
| effective Oct 23, 2026 | OpenAI | The OpenAI API keeps running. The gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4-1106-preview, gpt-4.1-nano, gpt-4o-2024-05-13, gpt-image-1, o1, o1-pro, o3-mini and o4-mini models, with their fine-tuned versions, can be called until October 23, 2026, when OpenAI shuts down their access. OpenAI names a substitute for each model: gpt-3.5-turbo to gpt-5.6-terra; gpt-4, gpt-4-turbo, gpt-4-1106-preview and gpt-4o-2024-05-13 to gpt-5.6-sol; gpt-4.1-nano to gpt-5.6-luna; gpt-image-1 to gpt-image-2.5-sunburst or gpt-image-2.5-flare; o1 to gpt-5.6-sol; o1-pro to gpt-5.6-sol with reasoning.mode: pro; o3-mini to gpt-5.6-sol; o4-mini to gpt-5.6-terra. Source ↗ | MEDIUM |
| effective Oct 23, 2026 | Google Gemini API | The Gemini API keeps running. The Deep Research agent preview deep-research-pro-preview-12-2025 can be called until October 23, 2026, when Google shuts it down. Google names deep-research-preview-04-2026 as the replacement. Source ↗ | LOW |
| effective Oct 27, 2026 | NVIDIA NIM | NVIDIA's free API endpoints keep running, and GLM 5.3 Flash (z-ai/glm-5.3-flash) is not affected. GLM 5.3 (z-ai/glm-5.3) can be called on NVIDIA's free endpoints until October 27, 2026; NVIDIA's model page says it will be deprecated on 10/27/2026 and no longer supported after 10/28/2026. Source ↗ | LOW |
| effective Nov 17, 2026 | Google Gemini API | The Gemini API keeps running. The text-to-speech previews gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts and gemini-3.1-flash-tts-preview, and the Live API previews gemini-3.1-flash-live-preview and gemini-2.5-flash-native-audio-preview-12-2025, can be called until November 17, 2026, when Google shuts them down. Google names gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts as the replacement for each text-to-speech preview, and gemini-3.8-live for both Live API previews. Source ↗ | LOW |
| effective Nov 19, 2026 | Google Gemini API | The Gemini API keeps serving gemini-3.6-flash with no shutdown date. On Gemini Enterprise Agent Platform (formerly Vertex AI), gemini-3.6-flash can be called until November 19, 2026, when Google retires it. Google names gemini-3.8-flash as the replacement. Source ↗ | LOW |
| effective Nov 30, 2026 | OpenAI | The OpenAI API keeps running. Agent Builder, the Evals platform and the v1/prompts API can be used until November 30, 2026. Existing evals become read-only on October 31, 2026. ChatKit remains available. Migrate from Agent Builder to the Agents SDK or ChatGPT Workspace Agents. Migrate from Evals to Promptfoo. Move reusable prompt content into application code. Source ↗ | LOW |
| effective Nov 30, 2026 | Anthropic API | Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) retires from the Claude API on 2026-11-30, and requests to it fail after that date. Anthropic deprecated it on 2026-09-30 and names claude-sonnet-5-5 as the replacement. The Claude API and Anthropic's other models keep running. Amazon Bedrock and Google Cloud set their own retirement dates. Source ↗ | MEDIUM |
| effective Dec 1, 2026 | OpenAI | The OpenAI API keeps running. The gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest models can be called until December 1, 2026. OpenAI's substitutes are gpt-image-2.5-sunburst or gpt-image-2.5-flare. Source ↗ | LOW |
| effective Dec 11, 2026 | OpenAI | The OpenAI API keeps running. gpt-5-2025-08-07 is available until December 11, 2026, replaced by gpt-5.6-sol. gpt-5-mini-2025-08-07 is available until December 11, 2026, replaced by gpt-5.6-terra. gpt-5-nano-2025-08-07 is available until December 11, 2026, replaced by gpt-5.6-luna. gpt-5-pro-2025-10-06 is available until December 11, 2026, replaced by gpt-5.6-sol with reasoning.mode: pro. o3-2025-04-16 is available until December 11, 2026, replaced by gpt-5.6-sol. o3-pro-2025-06-10 is available until December 11, 2026, replaced by gpt-5.6-sol with reasoning.mode: pro. Source ↗ | MEDIUM |
| effective Jan 6, 2027 | OpenAI | The OpenAI API keeps running. The tts-1, tts-1-hd, gpt-4o-mini-tts-2025-03-20 and gpt-4o-mini-tts-2025-12-15 models can be called until January 6, 2027. OpenAI's replacement for each: gpt-realtime-2.1-mini. Source ↗ | LOW |
| effective Jan 14, 2027 | OpenAI | The OpenAI API keeps running. The gpt-4o (gpt-4o-2024-08-06 and gpt-4o-2024-11-20), gpt-4o-mini, gpt-4.1 and gpt-4.1-mini models, with their fine-tuned versions, can be called until January 14, 2027, when OpenAI shuts down their access. OpenAI names gpt-6-luna as the replacement for all four. Source ↗ | MEDIUM |
| effective Jan 20, 2027 | OpenAI | The OpenAI API keeps running. The gpt-realtime, gpt-4o-realtime, gpt-realtime-mini, gpt-4o-mini-realtime, gpt-audio, gpt-4o-audio, gpt-audio-mini, gpt-4o-mini-audio and gpt-4o-mini-transcribe-2025-03-20 models can be called until January 20, 2027. OpenAI's substitutes: gpt-realtime-2.1 for gpt-realtime and gpt-4o-realtime; gpt-realtime-2.1-mini for gpt-realtime-mini and gpt-4o-mini-realtime; gpt-audio-1.5 for gpt-audio, gpt-4o-audio, gpt-audio-mini and gpt-4o-mini-audio; gpt-4o-mini-transcribe-2025-12-15 for gpt-4o-mini-transcribe-2025-03-20. Source ↗ | LOW |
| effective Feb 26, 2027 | OpenAI | The OpenAI API keeps running. The whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize models can be called until February 26, 2027. OpenAI's substitute for each is gpt-live-transcribe or gpt-transcribe. Source ↗ | MEDIUM |
| effective Mar 15, 2027 | Google Gemini API | The Gemini API keeps running. Gemini 2.5 Flash Image (gemini-2.5-flash-image) can be called until March 15, 2027, when Google shuts it down. Google names gemini-3.1-flash-lite-image as the replacement. Source ↗ | LOW |
| effective Apr 1, 2027 | OpenAI | The OpenAI API keeps running. The gpt-5.3-codex, gpt-5.1 and gpt-5.4-nano models can be called until April 1, 2027. OpenAI's replacements: gpt-6-sol for gpt-5.3-codex and gpt-5.1; gpt-6-luna for gpt-5.4-nano. Source ↗ | MEDIUM |
| effective May 7, 2027 | Google Gemini API | The Gemini API keeps running. Gemini 3.1 Flash-Lite (gemini-3.1-flash-lite) can be called until May 7, 2027, when Google shuts it down. Google names gemini-3.5-flash-lite as the replacement. Source ↗ | LOW |
| effective May 14, 2028 | Google Gemini API | The Gemini API keeps running. The embedding model gemini-embedding-001 can be called until May 14, 2028, when Google shuts it down. Google names gemini-embedding-2 as the replacement. Source ↗ | LOW |
Groq (free, fast, no credit card) or OpenRouter (25+ free models, try different providers). Cloudflare Workers AI for 10K neurons/day without a credit card.
OpenAI GPT-6 Sol or GPT-6.1 Sol ($2/M in, $10/M out) — function calling, structured outputs; GPT-5.6 Terra costs $2/$12 and GPT-4o $2.50/$10. Claude Sonnet 5.5 ($2/$10/M) with a 1M-token context.
Claude Opus 5.5 ($4/$20/M), or Claude Fable 5.1 ($10/$50/M) for long-horizon agentic work. DeepSeek V4-Pro ($1.32/$3.96/M at peak, half off-peak), thinking mode on by default. OpenAI GPT-6 Astra ($10/$50/M) for the hardest end-to-end work.
deepseek-flash ($0.30/$1.20/M at peak) with 97-98% cache-hit discounts. OpenAI/Anthropic batch APIs at 50% off for async workloads.
deepseek-flash (1M context, $0.30/$1.20/M at peak). Claude Fable 5.1, Opus 5.5, Sonnet 5.5 or Sonnet 5 (1M context each).
Ollama (free, run locally) for development. Replicate or Baseten for hosted open-source models with dedicated infrastructure. Cloudflare Workers AI for edge inference.