aiapicostcalcUnit economicsPricesBrandsCompareBenchmarksSkills & PluginsFind my modelGuidesChanges

New AI models this week: releases, prices and API changes

Short answer: This feed lists 43 verified model, API, pricing, benchmark, skill and plugin updates, newest first (latest 18 November 2026). Every item links its source and date; the scanner checks the sources every three hours and only publishes what it could verify.

Last verified 2026-11-18 · Sources: src/updates.json (each row links its own source page and check date) · Report an error

By Shangyi Duan · Every item links its own source and check date

Latest verified updates

18 November 2026

Price change · GPT-6 Rosalind pricing

GPT-6 Rosalind moves to $10/$50 short context, $20/$75 long context

OpenAI's pricing page replaced the single $15/$75 GPT-6 Rosalind row with short-context rates of $10/$50 per 1M (cached $1) and long-context rates of $20/$75. The promotional pricing runs at least through December 31, 2026, and standalone FedRAMP endpoints are 10% above the corresponding global rates.

Source: platform.openai.com · checked 2026-11-18

Price change · DeepSeek Flash (V4.1) pricing

DeepSeek stops billing search and web fetch that a model call triggers

From 12:00 on November 18, 2026 (Beijing time, UTC+8) server-side search and web fetch that a model call triggers are no longer billed; only token usage is charged. Direct Search and Fetch API calls still cost $0.75 per 1,000 successful requests, alongside file parsing at $0.225 per 1,000 pages and the speech rates.

Source: api-docs.deepseek.com · checked 2026-11-18

17 November 2026

New model

Introducing GPT-6.2 Astra and Sol

GPT-6.2 Astra is listed at $8/$40 per 1M tokens (short context) and GPT-6.2 Sol at $2/$10, matching GPT-6.1 Sol, with Batch, Fast and long-context tiers on the official pricing page. Ultrafast costs 3 times the Standard rate through February 28, 2027 and 6 times from March 1, 2027.

Source: openai.com · checked 2026-11-17

13 November 2026

API change

DeepSeek opens web search, file input, OCR, ASR and TTS APIs

DeepSeek's official change log adds web search, multi-step search while thinking, file input, URL reading and a safety check to deepseek-flash, and ships standalone Search, Fetch, OCR, ASR and TTS APIs billed per use: $0.75 per 1,000 searches or fetches, $0.225 per 1,000 parsed or OCR pages, $0.45 per 10,000 TTS characters and $0.116 per ASR hour of audio.

Source: api-docs.deepseek.com · checked 2026-11-13

12 November 2026

New model

Google launches Gemini 4 Hydrogen for agentic workflows

Google's efficiency-focused Gemini 4 model brings a 1M-token context window; official pricing is $1.50 input and $7.50 output per 1M tokens ($0.075 cached), with batch at $0.75/$3.75.

Source: blog.google · checked 2026-11-12

Price change · GPT-5.6 Sol pricing

GPT-5.6 Sol promotional pricing is available until 2026-11-30

OpenAI's pricing page now states that GPT-5.6 Sol's promotional pricing is available until November 30, 2026, and that GPT-6 Astra for Law's promotional pricing is available at least through December 31, 2026. The current list rates are unchanged.

Source: platform.openai.com · checked 2026-11-12

11 November 2026

New model

Gemini Decision API: $0.10 per 1M input tokens, no output tokens

Google added Gemini Decision (gemini-decision-1), a multimodal decision model for binary, categorical and rubric evaluations through the new Decisions API: $0.10 per 1M input tokens, shared input billed once per request, and no output tokens.

Source: ai.google.dev · checked 2026-11-11

New model · Gemini 4 Carbon pricing

Gemini 4 Carbon ships at $2/$10 per 1M tokens, rising to $4/$20 on 2026-12-01

Google added Gemini 4 Carbon, its most capable reasoning model, to the Gemini API pricing page: $2 input / $10 output per 1M tokens through 2026-11-30, then $4/$20.

Source: ai.google.dev · checked 2026-11-11

10 November 2026

New model

Google Cloud launches Gemini for healthcare.

Google Cloud launches Gemini for healthcare to unify clinical data and automate workflows. Reduce administrative burnout.

Source: blog.google · checked 2026-11-10

9 November 2026

New model

Meta Muse Ember (muse-ember-1.0) priced at $0.10 input / $0.50 output per 1M

Meta's research blog lists muse-ember-1.0 at $0.10 per 1M input, $0.004 cached input and $0.50 per 1M output tokens on the Meta Model API, with a 1M-token context window and up to 512K output tokens. Checked 2026-11-10; not yet in the price tables because they cover six brands.

Source: research.meta.ai · checked 2026-11-09

Benchmark · Claude Fable 5.2 pricing

Claude Fable 5.2 tops the Artificial Analysis Intelligence Index at 65

Artificial Analysis lists Claude Fable 5.2 at 65 on its Intelligence Index, the highest score on the leaderboard (#1 of 74), with 101 tok/s output speed. Scores read 2026-11-10.

Source: artificialanalysis.ai · checked 2026-11-09

New model · Claude Fable 5.2 pricing

Claude Fable 5.2 and Mythos 5.2 priced at $8 input / $40 output per 1M

Anthropic's official pricing page lists Claude Fable 5.2 and Claude Mythos 5.2 at $8 per 1M input, $0.20 cached input and $40 per 1M output tokens, with fast mode at $16 / $80; both released November 9, 2026. Prices checked 2026-11-10.

Source: docs.anthropic.com · checked 2026-11-09

New model

Agnes 2.5 Pro Theta priced at $0.10 input / $0.30 output per 1M tokens

Agnes AI's official pricing doc lists agnes-2.5-pro-theta at $0.10 per 1M input, $0.01 cached input and $0.30 per 1M output; the model also has a benchmark profile on Artificial Analysis. Prices checked 2026-11-09.

Source: wiki.agnes-ai.com · checked 2026-11-09

8 November 2026

Benchmark

Thinking Machines' Inkling 2 Small joins Artificial Analysis at $0.30/$1.00 per 1M

Artificial Analysis lists Inkling 2 Small at $0.30 per 1M input and $1.00 per 1M output tokens, about $1.18 per evaluation task. Checked 2026-11-08; official per-token pricing page not verified, so it is not in our price tables yet.

Source: artificialanalysis.ai · checked 2026-11-08

Benchmark

Tencent Hy4 joins Artificial Analysis leaderboard at $0.83/$2.50 per 1M tokens

Artificial Analysis lists Tencent Hy4 at $0.83 per 1M input and $2.50 per 1M output tokens, about $1.03 per evaluation task. Checked 2026-11-08; no official per-token pricing page verified, so it stays out of our price tables.

Source: artificialanalysis.ai · checked 2026-11-08

Benchmark

Meta's Muse Wonder (high) joins Artificial Analysis leaderboard at $4/$20 per 1M

Artificial Analysis now profiles Muse Wonder (High) at $4.00 per 1M input and $20.00 per 1M output tokens, about $4.56 per evaluation task. Listing checked 2026-11-08; no official per-token pricing page published yet, so it is not in our price tables.

Source: artificialanalysis.ai · checked 2026-11-08

6 November 2026

API change

DeepSeek moves Flash to V4.1-Flash-1017 and prices Search & Fetch APIs

DeepSeek's pricing page now lists deepseek-flash as served by DeepSeek-V4.1-Flash-1017 (legacy deepseek-v4-flash names retired, still billed at the Flash price) and prices its standalone Search and Fetch APIs at $0.75 per 1,000 successful requests. Per-token Flash and V4 Pro rates are unchanged (checked 2026-11-06).

Source: api-docs.deepseek.com · checked 2026-11-06

5 November 2026

New model · GPT-6 Cyber pricing

OpenAI releases GPT-6 Cyber for authorized security testing

OpenAI's GPT-6 Cyber is a Daybreak Red model for vulnerability investigation and fix verification in authorized environments: $20 input / $100 output / $2 cached per 1M tokens, 1.05M context and 128K max output, with no Batch API. Access requires approval (checked 2026-11-06).

Source: platform.openai.com · checked 2026-11-05

Price change

OpenAI lists Rosalind and Cyber model prices

OpenAI's official pricing page now publishes per-token prices for the Rosalind life-sciences family (gpt-6-rosalind $15/$75, gpt-rosalind-research and gpt-rosalind-discovery $5/$25) and the GPT-5.6 Cyber variant ($12.50/$75). All require restricted access; the site now tracks these models (checked 2026-11-06).

Source: platform.openai.com · checked 2026-11-05

API change

Mistral Studio launches: build enterprise agents, workflows and apps

Mistral announced Studio, its enterprise AI platform for building agents, workflows and apps on Mistral models, deployable in the cloud, a VPC or on-premises; the post positions Mistral Large 4 (Le Chonk) as the model behind it. Read from mistral.ai on 2026-11-05.

Source: mistral.ai · checked 2026-11-05

New model

Google launches Gemini Omni 1.2 Flash for longer video clips

Google's Gemini Omni 1.2 Flash generates video clips up to 15 seconds at lower pricing (about $0.70 per 10-second clip, $1.04 per 15-second clip; per-second rates for 360p/1080p/4K). It is billed per second rather than per token, so it is listed here as a release and not in the per-token price table (checked 2026-11-06).

Source: blog.google · checked 2026-11-05

4 November 2026

New model

Ricursive releases Ripple-65B Preview, open-weight SOTA on agentic coding

Ricursive Intelligence launched Ripple-65B Preview, an open-weight 65B dense model with up to 1M context and a claimed 80.52 on TerminalBench 2.1 (US open-weight SOTA). Available on Hugging Face, Together AI, GMI Cloud and OpenRouter; no official per-token list price has been published yet, so hosted rates are third-party prices.

Source: www.ricursive.com · checked 2026-11-04

Price change · GPT-5.6 Sol pricing

OpenAI promotional API pricing through November 21, 2026 (superseded)

OpenAI's official pricing page (checked 2026-11-04) listed promotional pricing as available at least through November 21, 2026 for selected models. OpenAI's current pricing page states promotional pricing is available until November 30, 2026 (verified 2026-11-12; see the 2026-11-12 update).

Source: platform.openai.com · checked 2026-11-04

3 November 2026

Benchmark

StateBench: how language models track state over long documents

AI2 releases StateBench, a new benchmark that tests whether language models correctly track state changes across long documents.

Source: huggingface.co · checked 2026-11-03

Benchmark

CWE-bench v2: more coverage, quality and transparency

The CWE-bench team publishes v2 of its vulnerability benchmark with broader coverage, higher quality data and more transparent methodology.

Source: huggingface.co · checked 2026-11-03

2 November 2026

Plugin

Qodo brings context into AI code review with Codex

Qodo integrates company knowledge and coding standards into Codex so AI code review can catch P0 issues before a pull request is opened.

Source: openai.com · checked 2026-11-02

Price change

Gemini 2.5 Pro and 2.5 Flash deprecated; API requests reroute on December 1, 2026

Google's official changelog marks gemini-2.5-pro and gemini-2.5-flash deprecated: from December 1, 2026, requests route automatically to gemini-3.8-flash ($0.75/$3.75 per 1M tokens) and gemini-3.5-flash-lite ($0.30/$2.50) at those models' rates.

Source: ai.google.dev · checked 2026-11-02

29 October 2026

New model · Mistral Large 4 pricing

Mistral opens the weights of Mistral Large 4 (Le Chonk)

Mistral publishes the open weights of Mistral Large 4, a 1T-parameter multimodal model it calls the best open-weight model for coding.

Source: mistral.ai · checked 2026-10-29

New model

Introducing AnySpeechRecognizer and AnyDecisionModel: Open Models for Speech and Decisions

Update from Hugging Face Blog (RSS).

Source: huggingface.co · checked 2026-10-29

27 October 2026

New model

Liquid AI releases LFM3 for specialized multimodal models

Liquid AI publishes LFM3, a new model family for building specialized multimodal models.

Source: huggingface.co · checked 2026-10-27

26 October 2026

New model · GPT-6.1 Terra pricing

Introducing GPT-6.1 Terra

OpenAI releases GPT-6.1 Terra, positioned for fast and capable computer use at scale.

Source: openai.com · checked 2026-10-26

22 October 2026

API change

Build more responsive apps with the Decisions API

OpenAI introduces the Decisions API for fast, focused, low-cost decisions inside applications.

Source: openai.com · checked 2026-10-22

Skill

Anthropic donates Agent Skills to Agentic AI Foundation

Anthropic contributes Agent Skills, its open standard for reusable agent knowledge and workflows, to the Agentic AI Foundation.

Source: www.anthropic.com · checked 2026-10-22

21 October 2026

Price change

Vocabulary specialization can cut SLM inference cost by 77%

A Hugging Face blog post reports that specializing the vocabulary of small language models cut inference cost by 77% in real applications.

Source: huggingface.co · checked 2026-10-21

API change

OpenAI models are coming to Google Cloud

OpenAI and Google Cloud bring frontier models and Codex workflows to Gemini's Model Garden, available through Google Cloud Marketplace.

Source: openai.com · checked 2026-10-21

20 October 2026

New model

Voxtral 3: real-time speech models for 50+ languages

Mistral releases Voxtral 3, a voice agent SDK with real-time speech in 50+ languages, transcription and emotional TTS.

Source: mistral.ai · checked 2026-10-20

17 October 2026

New model · DeepSeek Flash (V4.1) pricing

DeepSeek-V4.1-Flash-1017 update

DeepSeek upgrades the Flash API model to DeepSeek-V4.1-Flash-1017 while keeping the deepseek-flash model name and pricing.

Source: api-docs.deepseek.com · checked 2026-10-17

API change

DeepSeek opens its in-house search API

DeepSeek exposes its in-house search engine through the API, adding web search and fetch powered by its own index.

Source: api-docs.deepseek.com · checked 2026-10-17

16 October 2026

New model

Welcome Beam: a 501B open model you can own and scale

Reflection releases Beam, a 501B-parameter open model that teams can own, customize and scale.

Source: huggingface.co · checked 2026-10-16

Plugin

Stack Overflow for Agents triples signups with ChatGPT plugin

Two engineers ship Stack Overflow for Agents with Codex, and its ChatGPT plugin helps drive up to 10x growth in agent creation.

Source: openai.com · checked 2026-10-16

15 October 2026

New model

Meta open-weights Muse Spark Open, a frontier multimodal MoE

Meta publishes Muse Spark Open, an open-weight frontier multimodal mixture-of-experts model.

Source: huggingface.co · checked 2026-10-15

New model

Introducing Timestral

Mistral introduces Timestral, a state-of-the-art probabilistic forecasting model family in four sizes from 5M to 209M parameters.

Source: mistral.ai · checked 2026-10-15

New model

Apertus 1.5 Release Notes

Update from Hugging Face Blog (RSS).

Source: huggingface.co · checked 2026-10-15

43 items across 6 categories: 21 New model, 7 Price change, 6 API change, 6 Benchmark, 2 Plugin, 1 Skill. Feed last reviewed 2026-11-18; source pages are linked on every row.

How this feed is checked

Every three hours

A scanner runs on a three-hour schedule against the sources listed below: provider news and changelog pages, benchmark leaderboards and AI tool registries. It stores a fingerprint of every item it has seen, so a page that has not changed produces no entry and a changed one is flagged for review.

What gets published

An item appears only when the change can be read on its own source page and dated. Model launches, API changes, price changes, benchmark releases, new skills and new plugins all qualify. Each row links to the page that was read, with the date it was checked.

What does not get published

Rumours, screenshots without a source, price changes read from a reseller or a blog, and benchmark claims that cannot be traced to a leaderboard. An empty run is reported as an empty run; the feed does not invent a slow news day.

Corrections

If a source page changes after we publish, the item stays but the next check adds a newer row and the site-wide check date moves. Report a wrong item through the corrections link in the header.

Which sources are watched

Category by category

CategoryLabel in the feedPublished items
New modelmodel21
API changeapi6
Price changepricing7
Benchmarkbenchmark6
Skillskill1
Pluginplugin2

Counts are the items currently published in each category, not the number of sources scanned. Source pages are linked on every row of the feed above.

What a good entry contains

A title, a date, one or two sentences of what changed, the source URL and — when the item touches a model we price — a link to that model's price page. Anything short of that stays out of the feed.

Subscribe to changes

One short email when a model, API or price changes — no digest of every release. The signup form is at the bottom of this page; the same form is used for price-change alerts on every calculator page.

Related

Questions

How often is this page updated?

A scanner checks the watched sources every three hours and appends verified changes. The feed is rebuilt from src/updates.json, so a new item appears on the next build after it is verified; the latest check date is shown at the top of the feed.

Where do the items come from?

Provider news and changelog pages, benchmark leaderboards, and AI tool and plugin registries. Every row links the page that was read and the date it was checked; an item without a working source link is not published.

Why does the feed not include every release?

Because a release note is not automatically a change worth an entry. Items are published when the change is confirmed on its own source page; rumours, reseller prices and untraceable benchmark claims stay out.

What kinds of changes are tracked?

New models, API changes, price changes, new benchmark results, new skills and new plugins — the six categories labelled on every row. Items that touch a model we price carry a link to that model's price page.

Can I get updates by email?

Yes — the subscribe form at the bottom of this page sends one short email when a model or price changes. There is no daily digest and unsubscribing removes the address.

Next step → Compare the newest models on price

Price data last checked 2026-11-18 against official sources · how we verify

Price-change alerts

One short email when a provider changes a price or a promo ends. No spam, unsubscribe any time.