New AI models in 2026: the major API releases and what they cost

This page logs the major AI model releases of 2026 on the first-party APIs of eight providers: OpenAI, Anthropic, Google (Gemini API), xAI, DeepSeek, Mistral, Alibaba (Qwen on Model Studio) and Meta (Meta Model API). Each entry carries the model’s list price, and a separate table lists the price changes and shutdowns already scheduled. The log starts in January 2026 and is updated monthly. This version is as of 9 November 2026.

What counts: new text and reasoning models, and price changes to them, as each vendor’s own release notes, pricing pages and deprecation pages record them. Not logged: image, video, audio and embedding models; open-weight releases with no first-party API price; dated snapshot refreshes of an existing model (except qwen3-max-2026-01-23, which merged Qwen3-Max’s thinking and non-thinking modes); other labs’ models resold on a vendor’s platform; and specialist models sold only through an approval program, which the month notes mention instead. Alibaba dates are the Singapore-region listing dates on Model Studio’s model lifecycle page. xAI publishes only the month for Grok 4.20 and Grok Build 0.1, and does not date Grok 4.3 at all, so Grok 4.3 is not logged. Meta’s rows start in July, when it opened the Meta Model API in public preview; earlier 2026 changes to Meta’s Llama API are not covered.

Prices are each vendor’s list price in USD per million tokens as of 9 November 2026, which can differ from the launch price; the notes say when it did. A dash means the model is retired or no longer on the vendor’s price list. For live numbers, use the pricing explorer. The price calendar lists every dated price change and shutdown, including the Amazon Bedrock and Azure routes, and the calculator prices your own token mix.

The latest releases: November 2026 so far, and October

Also in November: OpenAI released GPT-6-Cyber on 5 November at $20.00 / $100.00; its model page says “Your organization needs Daybreak Red access and approval to use GPT-6-Cyber.” Google deprecated Gemini 2.5 Pro and Gemini 2.5 Flash on 2 November; from 1 December their requests are rerouted and billed at the new model’s rates (see the schedule below). Anthropic’s, xAI’s, DeepSeek’s and Alibaba Model Studio’s release notes list no new text models between 1 and 9 November.

• ◦Also in October, prices: Anthropic cut Claude Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens on 7 October. OpenAI added a 12-hour prompt cache on 20 October; 12-hour cache writes cost 2.5× the input rate.
• ◦Also in October, launches: Mistral Large 4 entered public preview on 6 October with launch pricing of “50% off for 2 weeks”. OpenAI released GPT-6 Rosalind on 30 October at $15.00 / $75.00; OpenAI says “Access is available to organizations approved for Rosalind Research through the trusted-access program.” DeepSeek upgraded deepseek-flash to V4.1-Flash-1017 on 17 October; DeepSeek says “The API calling method and pricing remain unchanged”.
• ◦Also in October, retirements: OpenAI shut down o1, o1-pro, o3-mini, o4-mini, GPT-4 Turbo, gpt-4.1-nano and other legacy snapshots on 23 October. Google deprecated Gemini 3.7 Flash and Gemini 3.5 Flash on 8 October and now routes their requests to Gemini 3.8 Flash and Gemini 3.6 Flash. Mistral deprecated Mistral Medium 3.5 and Mistral Large 3 on 31 October; both retire on 30 April 2027, with Large 4 as the replacement.

The 2026 release log, month by month

November and October are in the table above, and the rest of the year follows, newest first. The October and September deep dives are linked above; this hub summarizes, and the monthly posts carry the analysis.

September 2026: GPT-6 Astra, Claude Fable 5.1, Grok 4.7 and the GPT-6 family

Also in September: Alibaba listed a new Qwen3.8-Max snapshot (qwen3.8-max-0902) on 2 September at the same $2.00 / $6.00. OpenAI made GPT-Rosalind generally available on 8 September through its life-sciences trusted-access program, at $5.00 / $25.00. Anthropic deprecated Claude Sonnet 4.5 on 30 September, with retirement set for 30 November.

August 2026: Qwen3.8-Max, Grok 4.6, DeepSeek V4-Pro GA and a GPT-5.6 Sol price cut

Also in August: Anthropic retired Claude Opus 4.1 on the Claude API on 5 August. OpenAI made GPT-5.6 Cyber available on 7 August through Daybreak Red, which it describes as “separately approved access”, at $12.50 / $75.00. DeepSeek released an experimental vision model, deepseek-v4-flash-vision-exp, on 21 August; it has since been retired, and its requests are served by V4.1-Flash at Flash rates.

July 2026: GPT-5.6, Claude Opus 5, Gemini 3.6 Flash, Grok 4.5 and the Meta Model API

Also in July: OpenAI shut down gpt-5-codex, the GPT-5.1 Codex models, gpt-5.2-codex and other legacy snapshots on 23 July. DeepSeek moved its V4-Flash API to an official release, “now in public beta”, on 31 July; V4-Flash was retired on 10 September when V4.1-Flash replaced it.

June 2026: Claude Fable 5, Claude Sonnet 5 and Qwen3.7-Plus

Also in June: Google shut down Gemini 2.0 Flash and Gemini 2.0 Flash-Lite on 1 June. On 11 June, OpenAI announced the 11 December shutdown of its GPT-5 and o3 snapshots. Anthropic retired Claude Sonnet 4 and Claude Opus 4 on the Claude API on 15 June. Claude Fable 5 and Mythos 5 use the tokenizer introduced with Claude Opus 4.7, which Anthropic says produces roughly 30% more tokens for the same text than models before Opus 4.7.

May 2026: Claude Opus 4.8, Qwen3.7-Max, Gemini 3.5 Flash and an xAI retirement

Also in May: xAI retired grok-3, grok-4-0709, the grok-4-fast and grok-4-1-fast models and grok-code-fast-1 on 15 May. The slugs still resolve. Requests to most of them now go to grok-4.3 and bill at its $1.25 / $2.50, not the original model’s rates; grok-code-fast-1 now goes to grok-build-0.1. Google shut down the Gemini 3.1 Flash-Lite preview on 25 May.

April 2026: GPT-5.5, DeepSeek V4, Claude Opus 4.7, Qwen3.6 and Mistral Medium 3.5

Also in April: Anthropic retired Claude Haiku 3 on 20 April. Google added the Gemma 4 open models to the Gemini API on 2 April; they have no paid-tier price.

March 2026: GPT-5.4, Mistral Small 4 and Grok 4.20

Also in March: Google shut down Gemini 3 Pro Preview on 9 March and pointed gemini-3-pro-preview at Gemini 3.1 Pro Preview. On 13 March, Anthropic took the 1M-token context window out of beta for Claude Opus 4.6 and Sonnet 4.6, at standard pricing.

February 2026: Claude Opus 4.6, Claude Sonnet 4.6, Gemini 3.1 Pro Preview and Qwen3.5

Also in February: Anthropic retired Claude Sonnet 3.7 and Claude Haiku 3.5 on the Claude API on 19 February.

January 2026: gpt-5.2-codex and a Qwen3-Max snapshot

Also in January: Anthropic retired Claude Opus 3 on 5 January. On 27 January, Mistral moved Devstral 2 to paid API access; Devstral 2 has since been deprecated and retired on 31 July. Google’s, xAI’s and DeepSeek’s release notes list no new text models in January.

Upcoming price changes and shutdowns

These changes are already on the vendors’ own pricing and deprecation pages. Two of them raise prices without any change to your code: Gemini 4 Argon on 1 December and the Gemini Flash introductory rates on 1 January. Google’s 2.5 reroutes also change what you pay, because rerouted requests bill at the new model’s rates.

OpenAI’s audio and transcription shutdowns on 6 January, 20 January and 26 February 2027, and the Bedrock and Azure retirements, are on the price calendar.

Current flagship prices by provider

The top general-purpose API models from each provider, as of 9 November 2026. For the workhorse, reasoning and budget tiers, and for the caveats that reorder this table, read the full comparison.

Thresholds change these numbers. GPT-6 Astra bills $20.00 / $75.00 once a prompt passes 272K tokens, applied to the whole request. Grok 4.7 doubles every rate once a prompt reaches 200K tokens. DeepSeek V4 Pro costs $1.32 / $3.96 at peak hours. Gemini 4 Argon doubles on 1 December.

A release changes your bill only when your traffic moves to it, or when the model you already run changes price or gets rerouted. Both show up in per-model spend before they show up on an invoice.

What are the newest AI models?

As of 9 November 2026, the newest text models on the first-party APIs tracked here are Mistral Large 4, which Mistral’s model card lists as GA with a release date of 5 November; OpenAI’s GPT-6.1 Terra (26 October); Google’s Gemini 4 Argon (13 October); Meta’s Muse Wonder, in preview (12 October); and Anthropic’s Claude Haiku 5.5 (7 October). The flagship table above lists each provider’s top models today.

Which new AI model is cheapest?

Among 2026 releases still on sale, Alibaba’s Qwen3.7-Flash has the lowest list rates: $0.03 input and $0.13 output per million tokens, for prompts up to 32K tokens at the Singapore list price. Next come Qwen3.5-Flash at $0.10 / $0.40, Claude Haiku 5.5 and GPT-6 Luna at $0.10 / $0.50, Qwen3.8-Flash at $0.15 / $0.47, then Mistral Small 4 and DeepSeek V4.1-Flash (off-peak) at $0.15 / $0.60. Haiku 5.5 rises to $0.50 / $2.50 on prompts over 100K tokens, and DeepSeek doubles at peak hours, so rank on your own input:output mix rather than the headline rate.

How often do AI API prices change?

More than once a month on average across these eight providers: between July and October 2026 they made at least five list-price changes and cancelled one scheduled increase. OpenAI cut GPT-5.6 Luna by 80% and Terra by 20% (30 July). DeepSeek announced peak/off-peak pricing (13 August), then lowered its Flash prices with V4.1-Flash (10 September). OpenAI put GPT-5.6 Sol on a promotional $4 / $20 (21 August), and Anthropic halved Claude Sonnet 5.5’s cache-read rate (7 October). Anthropic also cancelled Claude Sonnet 5’s scheduled rise to $3 / $15 and kept $2 / $10 as the standard price (10 August). Two scheduled increases come next: Gemini 4 Argon on 1 December, and the Gemini 3.8 and 3.6 Flash introductory rates on 1 January 2027.

When do old models shut down?

Each vendor publishes its own schedule, and the notice differs. OpenAI promises at least six months for generally available models, at least three for specialized variants, and much shorter notice, such as two weeks, for preview models. Anthropic gives at least 60 days’ notice for publicly released models. Google lists the earliest possible shutdown date, and for some deprecated Gemini models it reroutes requests to a newer model and bills that model’s rates instead of returning an error. The next first-party shutdowns are Claude Sonnet 4.5 on 30 November, OpenAI’s older image models and Google’s gemini-2.5-pro and gemini-2.5-flash (rerouted, not switched off) on 1 December, OpenAI’s GPT-5 and o3 snapshots on 11 December, and GPT-4o and GPT-4.1 on 14 January 2027.

How current is this list?

It is checked against each vendor’s release notes, pricing pages and deprecation pages and updated monthly; this version is as of 9 November 2026. Release dates are the vendor’s own: for Alibaba, the Singapore-region listing date, and for Grok 4.20 and Grok Build 0.1 only the month, because xAI publishes no day. Prices can move between updates, so check the live pricing explorer before you commit a budget.