What do LLM APIs cost per million tokens in October 2026?
On October 30, 2026 the current text models from OpenAI, Anthropic and Google list from $0.10 to $10 per million input tokens and $0.50 to $50 per million output tokens. The cheapest tier is GPT-6 Luna and Claude Haiku 5.5, both $0.10 and $0.50.
We read each provider's own pricing, deprecation and release-note pages on October 30, 2026 and recorded every current text model, every announced retirement and every scheduled price change in one dataset. The table lists the standard, pay-as-you-go price for short prompts in US dollars per million tokens. It is the information behind the cost section of our AI agent development page, set out in full so you can check the models your own agents use.
| Provider | Model (API ID) | Input per 1M | Output per 1M | Note from the pricing page |
|---|---|---|---|---|
| OpenAI | gpt-6-astra | $10.00 | $50.00 | Flagship |
| OpenAI | gpt-6.1-sol | $2.00 | $10.00 | Released September 29, 2026 |
| OpenAI | gpt-6-sol | $2.00 | $10.00 | Released September 22, 2026 |
| OpenAI | gpt-6.1-terra | $1.00 | $5.00 | Released October 26, 2026 |
| OpenAI | gpt-6-luna | $0.10 | $0.50 | Released September 22, 2026 |
| OpenAI | gpt-5.6-sol | $4.00 | $20.00 | Promotional price, available at least through November 21, 2026 |
| OpenAI | gpt-5.6-terra | $2.00 | $12.00 | |
| OpenAI | gpt-5.6-luna | $0.20 | $1.20 | |
| Anthropic | Claude Fable 5.1 | $10.00 | $50.00 | |
| Anthropic | Claude Opus 5.5 | $4.00 | $20.00 | Launched September 22, 2026 |
| Anthropic | Claude Opus 5 | $5.00 | $25.00 | |
| Anthropic | Claude Sonnet 5.5 | $2.00 | $10.00 | Launched September 28, 2026 |
| Anthropic | Claude Sonnet 5 | $2.00 | $10.00 | Introductory price made standard; planned rise cancelled |
| Anthropic | Claude Haiku 5.5 | $0.10 | $0.50 | Prompts up to 100,000 tokens; $0.50 and $2.50 above that |
| gemini-4-argon | $2.00 | $10.00 | Rises to $4.00 and $20.00 on December 1, 2026 | |
| gemini-3.1-pro-preview | $2.00 | $12.00 | Prompts up to 200,000 tokens | |
| gemini-3.8-flash | $0.75 | $3.75 | Rises to $1.50 and $7.50 on January 1, 2027 | |
| gemini-3.6-flash | $0.75 | $3.75 | Rises to $1.50 and $7.50 on January 1, 2027 | |
| gemini-3.5-flash-lite | $0.30 | $2.50 | ||
| gemini-3.1-flash-lite | $0.25 | $1.50 | Earliest shutdown May 7, 2027 |
Three things the per-token price does not show. Reasoning models bill their thinking as output: OpenAI says reasoning tokens "are billed as output tokens," and Google's Gemini 4 Argon output price is listed as "including thinking tokens." Batch processing costs half on all three providers' pages. And Anthropic notes that Claude 4.7 and later models use a tokenizer that "produces approximately 30% more tokens for the same text," so equal prices per token do not mean equal prices per task.
Which AI models are being retired, and when?
Fifteen text models from OpenAI, Anthropic and Google have shutdown dates between November 30, 2026 and May 7, 2027. Most are OpenAI's: the GPT-5 family and o3 on December 11, 2026, then GPT-4o and GPT-4.1 models on January 14, 2027.
Shutdowns and price changes ahead
| Shutdown date | Provider | Models | Recommended replacement | Announced |
|---|---|---|---|---|
| November 30, 2026 | Anthropic | claude-sonnet-4-5-20250929 | claude-sonnet-5-5 | September 30, 2026 |
| December 11, 2026 | OpenAI | gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, o3-pro | gpt-5.6-sol, gpt-5.6-terra or gpt-5.6-luna | June 11, 2026 |
| January 14, 2027 | OpenAI | gpt-4o (2024-08-06 and 2024-11-20), gpt-4o-mini, gpt-4.1, gpt-4.1-mini | gpt-6-luna | October 14, 2026 |
| April 1, 2027 | OpenAI | gpt-5.1, gpt-5.4-nano, gpt-5.3-codex | gpt-6-sol or gpt-6-luna | October 1, 2026 |
| May 7, 2027 (earliest) | gemini-3.1-flash-lite | gemini-3.5-flash-lite | Listed on the deprecations page |
Fine-tuned versions of the OpenAI models "follow the same shutdown schedule as their underlying base models." Google's dates are the "earliest possible dates on which a model might be retired." Two of Google's Flash models, gemini-3.7-flash and gemini-3.5-flash, were deprecated on October 8, 2026 without a shutdown: requests to them are "automatically routed" to gemini-3.8-flash and gemini-3.6-flash. Your code keeps working, but on a different model than the one you tested.
OpenAI is also closing three platform tools on November 30, 2026: Agent Builder, the Evals platform (read-only from October 31) and reusable prompt objects. Teams that built agents or tests in those tools need to move them. We covered the most urgent group in OpenAI retires GPT-4o and GPT-4.1 on January 14, 2027.
How much notice do OpenAI, Anthropic and Google give before retiring a model?
Anthropic gave 60 to 62 days on all four 2026 retirements. OpenAI gave 182 to 184 days for most general models, but 92 days for GPT-4o and GPT-4.1. Google gave 11 to 103 days.
Days of notice before a model shutdown, 2026
| Provider | Stated policy | Notices measured in 2026 | Days of notice |
|---|---|---|---|
| Anthropic | "At least 60 days' notice before model retirement for publicly released models" | Sonnet 4.5; Opus 4.1; Sonnet 4 and Opus 4; Haiku 3 | 61; 61; 62; 60 |
| OpenAI | At least 6 months for generally available models, 3 months for specialized variants, possibly 2 weeks for previews | GPT-5 and o3 group; GPT-5.1 group; legacy group shut down October 23 | 183; 182; 184 |
| OpenAI | (same) | GPT-4o and GPT-4.1 group; chat-latest snapshots; July legacy group; GPT-5.4-Cyber | 92; 94; 92; 20 |
| Shutdown dates are the earliest possible; exact dates communicated "with advance notice" | Gemini 2.0 Flash and Flash-Lite; 3 Pro Preview; 2.5 Flash-Lite Preview; 3.1 Flash-Lite Preview | 103; 11; 21; 18 |
Days are counted from the announcement date on the provider's page to the shutdown date. One result stands out. OpenAI's page lists a minimum of 6 months for generally available models "unless safety or compliance concerns require a faster timeline," but the October 14, 2026 notice for GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini gives 92 days. The page does not say why. Anthropic is the most predictable: every 2026 notice sat within two days of its 60-day minimum.
Does the replacement model cost more or less than the one being retired?
It depends on the pair. GPT-4o to GPT-6 Luna cuts our sample workload from $62.00 to $2.60 per 1,000 tasks, and Claude Sonnet 4.5 to 5.5 cuts it by a third. GPT-5 mini to GPT-5.6 Terra raises it from $7.40 to $54.40.
Cost per 1,000 agent tasks, before and after
| Retiring model | Recommended replacement | Old price (in / out) | New price (in / out) | Workload per 1,000 tasks | Change |
|---|---|---|---|---|---|
| gpt-4o | gpt-6-luna | $2.50 / $10.00 | $0.10 / $0.50 | $62.00 to $2.60 | -96% |
| gpt-4.1 | gpt-6-luna | $2.00 / $8.00 | $0.10 / $0.50 | $49.60 to $2.60 | -95% |
| gpt-4.1-mini | gpt-6-luna | $0.40 / $1.60 | $0.10 / $0.50 | $9.92 to $2.60 | -74% |
| gpt-4o-mini | gpt-6-luna | $0.15 / $0.60 | $0.10 / $0.50 | $3.72 to $2.60 | -30% |
| gpt-5.4-nano | gpt-6-luna | $0.20 / $1.25 | $0.10 / $0.50 | $5.50 to $2.60 | -53% |
| claude-sonnet-4-5 | claude-sonnet-5-5 | $3.00 / $15.00 | $2.00 / $10.00 | $78.00 to $52.00 | -33% |
| gpt-5.1 | gpt-6-sol | $1.25 / $10.00 | $2.00 / $10.00 | $37.00 to $52.00 | +41% |
| gemini-3.1-flash-lite | gemini-3.5-flash-lite | $0.25 / $1.50 | $0.30 / $2.50 | $6.80 to $9.00 | +32% |
| o3 | gpt-5.6-sol | $2.00 / $8.00 | $4.00 / $20.00 | $49.60 to $104.00 | +110% |
| gpt-5 | gpt-5.6-sol | $1.25 / $10.00 | $4.00 / $20.00 | $37.00 to $104.00 | +181% |
| gpt-5-nano | gpt-5.6-luna | $0.05 / $0.40 | $0.20 / $1.20 | $1.48 to $5.44 | +268% |
| gpt-5-mini | gpt-5.6-terra | $0.25 / $2.00 | $2.00 / $12.00 | $7.40 to $54.40 | +635% |
The workload is the CRM-action agent from our money page: 20 million input and 1.2 million output tokens per 1,000 tasks, at standard prices with no caching. Arithmetic: (20 x input price) + (1.2 x output price). The gpt-5.6-sol price is OpenAI's promotional rate, so those two rows could rise after November 21, 2026.
The recommended replacement is not the only choice. OpenAI recommends gpt-5.6-terra for gpt-5-mini, but its own price table also lists gpt-6-luna at $0.10 and $0.50. Which one keeps your agent's quality is a question only your test set can answer, which is why we rerun it on two or three candidates before any switch, using the method in how to test an AI agent. For new builds we start on current models with no announced retirement, picked to fit each client's requirements.
Which LLM API prices are scheduled to change?
Google has published two increases: Gemini 4 Argon doubles to $4 and $20 per million tokens on December 1, 2026, and Gemini 3.8 and 3.6 Flash double to $1.50 and $7.50 on January 1, 2027. OpenAI's GPT-5.6 Sol promotion runs to at least November 21.
| Date | Provider | Model | Change |
|---|---|---|---|
| October 7, 2026 (done) | Anthropic | Claude Sonnet 5.5 | Cache reads cut from $0.20 to $0.10 per million tokens |
| At least to November 21, 2026 | OpenAI | gpt-5.6-sol | Promotional $4 / $20; OpenAI's changelog describes it as 20% lower input and 33% lower output pricing |
| December 1, 2026 | gemini-4-argon | $2 / $10 rises to $4 / $20; caching $0.10 to $0.20 | |
| January 1, 2027 | gemini-3.8-flash, gemini-3.6-flash | $0.75 / $3.75 rises to $1.50 / $7.50 |
Anthropic went the other way on one model: the $2 and $10 introductory price for Claude Sonnet 5 "is now the standard price," and "the previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur."
How did we build this dataset?
We read nine primary pages from the three providers on October 30, 2026: each one's pricing page, deprecations page and release notes. We kept text-generation models on the providers' own APIs and counted notice days from each announcement date to its shutdown date.
- Sources. OpenAI pricing, deprecations and changelog; Anthropic pricing, model deprecations and release notes; Gemini API pricing, deprecations and release notes. All links are in the sources list.
- Scope. Text and reasoning models only. We left out image, video, audio, speech, transcription, embedding and robotics models, and managed agents. Prices are standard tier, short prompts, USD per million tokens, before batch, cache or regional uplifts.
- Notice days. Calendar days from the announcement date the provider's page gives to the shutdown date. For Google we took announcement dates from the release notes, because the deprecations page lists release and shutdown dates only.
- Workload cost. One assumed workload, the same one used on our AI agent development page, so the figures can be compared. It assumes equal token counts on old and new models.
One more finding came out of the check. OpenAI's standard price table on October 30 still lists ten rows for models its deprecations page says were shut down on October 23, 2026, including gpt-3.5-turbo, gpt-4-0613, o1, o3-mini and o4-mini. A price on the pricing page is not proof that a model still runs. The deprecations page is the source to trust for status.
What does this data prove, and what does it not?
It proves what each provider published on October 30, 2026: the prices, the dates and the notice it gave. It does not show model quality, how many tokens a model uses for the same task, or whether a provider will change a date.
- It shows that every provider retires models on a schedule, that notice ranged from 11 to 184 days in 2026, and that a forced move can cut or multiply your token bill.
- It does not show quality. A cheaper replacement may need more steps or a stronger reasoning setting to match the old model on your job.
- It does not show real token use. Reasoning models add output tokens, and newer Claude models count about 30% more tokens for the same text, so our workload figures are a floor for comparison, not a forecast.
- It can change. Google's shutdown dates are the earliest possible; promotional and scheduled prices end. Check the provider pages again before you budget.
What should a business running AI agents do with this?
List every model ID your agents and no-code AI steps use, check each against the shutdown table, and move the ones retiring first. Rerun your test cases on the replacement and one cheaper alternative, then switch before the date, not on it.
In practice: find the model names in your n8n, Make, Zapier or GoHighLevel AI steps and in any custom code. Anthropic suggests exporting usage by API key and model from the Console; OpenAI's models endpoint reports a shutdown_date field for each model. Then budget for the price changes above. If you want help with the move, our AI agent development team does this as part of support, and our AI agent adoption research covers why agent projects stall when nobody owns this upkeep. For agents built inside n8n, our n8n automation service covers model changes in workflows.
More questions about LLM API prices and model retirements
Which is the cheapest LLM API in October 2026?
Among the current models on these three providers' pages, GPT-6 Luna and Claude Haiku 5.5 share the lowest standard price at $0.10 per million input and $0.50 per million output tokens; Haiku 5.5 rises to $0.50 and $2.50 above 100,000 prompt tokens. Older models such as gpt-5-nano list lower prices but retire on December 11, 2026. Cheapest per token is not always cheapest per task.
What happens to your agent on the shutdown date?
Requests to the retired model fail. Anthropic's page says "requests to models past the retirement date will fail," and OpenAI says the model "will no longer be accessible." An agent or automation step still pointing at that model ID stops working until someone changes it. Google's auto-routed Flash models are the exception: they keep answering, on a newer model.
How do you find out which models your automations use?
Check the model field in every AI step in your automation tools and the model string in any code. Then cross-check with the provider: Anthropic's Console usage export breaks usage down by API key and model, and OpenAI's List models endpoint reports each model's shutdown_date. Any model ID that appears in the shutdown table above needs a plan.
Is a deprecated model still safe to use until shutdown?
It still works, but plan to leave early. Anthropic warns that "deprecated models are likely to be less reliable than active models," and OpenAI advises against relying on preview models for business-critical work because they can be retired with as little as two weeks' notice. Use the remaining time to test the replacement.
How often will this tracker be updated?
This page reflects the providers' pages as read on October 30, 2026, and every figure carries that date. Providers add models and notices every few weeks, so check the source pages linked below before making a budget or migration decision, and treat any date here as a snapshot.
Sources and further reading
- OpenAI API pricing, checked October 30, 2026
- OpenAI deprecations (notice periods, all 2026 notices), checked October 30, 2026
- OpenAI API changelog (GPT-6 and GPT-6.1 releases, GPT-5.6 Sol promotional price), checked October 30, 2026
- OpenAI, Reasoning models (reasoning tokens billed as output), checked October 30, 2026
- Anthropic Claude pricing (model prices, tokenizer note, Sonnet 5 footnote), checked October 30, 2026
- Anthropic model deprecations, checked October 30, 2026
- Claude Platform release notes (Opus 5.5, Sonnet 5.5 and Haiku 5.5 launches; cache price cut), checked October 30, 2026
- Gemini API pricing (current and scheduled prices), checked October 30, 2026
- Gemini API deprecations, checked October 30, 2026
- Gemini API release notes (announcement dates, October 8 auto-routing), checked October 30, 2026
