What changed in 2026

gemini-4-argon became generally available on 13 October, the first Gemini 4 model in Google's API notes. The same day it reached the Gemini app for Google AI Pro and Ultra subscribers. Free accounts are not mentioned.

In October Databricks marked Knowledge Assistant and Supervisor Agent as legacy, 8 months after Supervisor Agent became generally available. It also published a guide to building agentsAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. in code with Agent Bricks.

From 12 October a fine-grained Hugging Face token can be given a contributor permission. It can create repositories and write to the ones it made, without write access to the other repositories around them. A script that uploads models no longer needs a token that can change everything.

On 8 October Google deprecated gemini-3.7-flash and gemini-3.5-flash. Calls to them now go to 3.8 Flash and 3.6 Flash. Your code keeps running, but on a different model, so check its answers.

Claude Haiku 5.5 came out on 7 October with a 1 million token context windowContext WindowEverything a model can see at once while it writes an answer, counted in tokens. When the window is full, something has to be left out.. It decides how long to think by itself, and returns an error if you send TemperatureTemperatureA setting that controls how much chance goes into each word a model picks. Low gives focused, predictable answers; high gives more varied ones., top_p, top_k, a thinking budget or a prefilled start of the answer. Code moving up from an older Haiku needs those removed.

From 7 October, reading cached text on Sonnet 5.5 costs $0.10 per million tokensTokenThe small piece of text a language model reads and writes, often part of a word. Prices, limits and the context window are all counted in tokens., down from $0.20. The same day, the Claude Max and Team plans began to include monthly API credits.

On 6 October OpenAI replaced its 5 usage tiers with 3: Build, Launch and Grow. Your account's rate limits come from its tier, so check where you landed.

Mistral released mistral-large-4 in public preview on 6 October, at 50% off for its first 2 weeks. Mistral says open weights are coming but has not released them yet.

xAI launched grok-4.7 in September with a 500,000 token context windowContext WindowEverything a model can see at once while it writes an answer, counted in tokens. When the window is full, something has to be left out., at the same prices as Grok 4.6. xAI's notes give the month only.

In September Google moved Gemini Live, Search Live and the voice features in Workspace to Gemini 3.8 Live, for every user. Google's roundup gives the month only.

Google released a Gemini app for Windows in September. Alt + Space opens it, and it can take on a task of several steps and work with your files from Gmail and Drive. Google's roundup gives the month only.

gpt-6-sol and gpt-6-luna came out on 22 September, and gpt-6.1-sol a week later, on 29 September. 3 new GPT-6 models in 4 weeks: pin the version your app was tested on.

Claude Sonnet 5.5 came out on 28 September. 2 days later Anthropic set 30 November 2026 as the retirement date for Claude Sonnet 4.5. If an app still runs on claude-sonnet-4-5-20250929, it has 2 months.

Claude Opus 5.5 came out on 22 September at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. It always thinks, and you cannot turn that off.

From 22 September, Transformers can load the compact quantizedQuantizationStoring a model's numbers with less precision, such as 4 bits instead of 16, so it needs far less memory and can run on smaller hardware, at some cost in accuracy. GGUF files made for llama.cpp, so a model sized for a laptop runs from the usual Python code. For now it needs the newest unreleased Transformers, and its fast path works only on Apple Silicon Macs and with Qwen models so far. Hugging Face still recommends llama.cpp when speed on your own machine matters most.

From 18 September, only people who already used the Gemini 2.5 models can call them. The models are not deprecated, but a new project cannot start on them.

DeepSeek released V4.1 Flash on 10 September, the smallest model in its new family, which reads images as well as text, and lowered its API prices. The old V4 Flash names now answer with V4.1 Flash for a while.

OpenAI released its Agents API in public beta on 10 September, and added computer use to it on 29 September. It is OpenAI's own place to run an AgentAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done., after it announced the end of Agent Builder in June.

gpt-6-astra, the first GPT-6 model in OpenAI's changelog, came out on 3 September. You can steer it in the middle of a turn, change how hard it thinks mid-conversation, and let its tool calls run while it keeps working.

gemini-3.8-flash became generally available on 2 September, 3 weeks after 3.7 Flash.

Claude Fable 5.1 and Claude Mythos 5.1 came out on 1 September at $10 per million input tokensTokenThe small piece of text a language model reads and writes, often part of a word. Prices, limits and the context window are all counted in tokens. and $50 per million output tokens. Like Fable 5, both need 30-day data retention and are not offered under zero data retention unless Anthropic expressly allows it. If your company needs zero retention, check this before you switch.

xAI launched grok-4.6 in August with a 500,000 token context windowContext WindowEverything a model can see at once while it writes an answer, counted in tokens. When the window is full, something has to be left out.. Its price doubles once a prompt passes 200,000 tokens, so a long prompt costs more per token than a short one. xAI's notes give the month only.

The Assistants API shut down on 26 August. OpenAI also set 26 February 2027 as the end for whisper-1 and the GPT-4o transcription models. If you turn speech into text with them, you have about 6 months to move.

From 21 August gpt-5.6-sol costs $4 per million input tokens and $20 per million output tokens. OpenAI calls it promotional pricing, held at least until 21 November 2026.

On 19 August Claude's computer use tool left beta, and a new browser use tool arrived. An AgentAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. now has a tool made for working in web pages. The Files and Skills APIs left beta the same day.

DeepSeek V4-Pro became generally available on 13 August. From 16 August, DeepSeek charges half price outside its peak hours. If a job can wait, run it off-peak. DeepSeek gives the hours in its notice for the change, linked from its news page.

gemini-3.7-flash became generally available on 13 August, with an introductory price that runs to 31 December 2026. Plan for the price to change after that.

On 10 August Anthropic cancelled the planned rise of Claude Sonnet 5 to $3 and $15 per million tokensTokenThe small piece of text a language model reads and writes, often part of a word. Prices, limits and the context window are all counted in tokens.. It stays at $2 per million input tokens and $10 per million output tokens.

Claude Opus 4.1 retired on 5 August, 2 months after Anthropic announced it. Calls to claude-opus-4-1-20250805 now fail.

DeepSeek opened the V4-Flash API in public beta on 31 July, under the name deepseek-v4-flash. The old names deepseek-chat and deepseek-reasoner were set to stop on 24 July, so use the new name.

xAI launched grok-4.5 in July at $2 per million input tokens and $6 per million output tokens, with 3 reasoning levels and high as the default. xAI's notes give the month only.

In July Databricks moved its Genie products to pay-as-you-go pricing and renamed Genie Spaces to Genie Agents. If your team uses Genie, its cost now follows its use.

On 30 July gpt-5.6-luna became 80% cheaper and gpt-5.6-terra 20% cheaper. Priority Processing became Fast mode, which runs gpt-5.6-sol up to 2.5 times faster at twice the price. Check whether a job you run on Sol could run on Luna.

Claude Opus 5 came out on 24 July at $5 per million input tokens and $25 per million output tokens, with a 1 million token window. It thinks by default, and you can turn thinking off only at effort high or below. Fast mode for Opus 4.7 ended the same day, so requests asking for it now fail.

From 22 July you can set a hard spend limit for an OpenAI organisation or a single project. Set one before you let an AgentAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. run on its own.

gemini-3.6-flash became generally available on 21 July at a lower price than 3.5 Flash, alongside gemini-3.5-flash-lite. Google also deprecated TemperatureTemperatureA setting that controls how much chance goes into each word a model picks. Low gives focused, predictable answers; high gives more varied ones., top_p and top_k. Stop sending them in new code.

GPT-5.6 came out on 9 July as a family of 3: gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna. The plain name gpt-5.6 answers with Sol. If you ask for gpt-5.6, know which of the 3 you are paying for.

Meta released Muse Spark 1.1 on 9 July, in Thinking mode in the Meta AI app, and opened a public preview of its Meta Model API so developers can call it. It works with a 1 million token context windowContext WindowEverything a model can see at once while it writes an answer, counted in tokens. When the window is full, something has to be left out..

Claude Sonnet 5 came out on 30 June at an introductory $2 per million input tokens and $10 per million output tokens, with a 1 million token window. Like the other Claude 5 models, it uses the new tokenizer, decides how long to think by itself, and rejects a non-default TemperatureTemperatureA setting that controls how much chance goes into each word a model picks. Low gives focused, predictable answers; high gives more varied ones..

In June Databricks renamed Vector Search to AI Search. If you search your own documents for RAGRAGA way to answer from your own documents. A search finds the passages that match the question, and they are added to the prompt before the model writes. on Databricks, look for it under the new name.

From 26 June, Sonnet and Haiku get the same rate limits as Opus, and the usage tiers shrank to 3: Start, Build and Scale. Busy apps on the cheaper models hit the limit less often.

Claude Sonnet 4 and Claude Opus 4 retired on 15 June. Calls to claude-sonnet-4-20250514 and claude-opus-4-20250514 now fail.

Claude Fable 5 and Claude Mythos 5 came out on 9 June, each with a 1 million token context windowContext WindowEverything a model can see at once while it writes an answer, counted in tokens. When the window is full, something has to be left out.. Their TokenizerTokenizerThe part that cuts text into the tokens a model reads, and turns the model's tokens back into text. Each model family has its own. cuts the same text into about 30% more tokensTokenThe small piece of text a language model reads and writes, often part of a word. Prices, limits and the context window are all counted in tokens., so the same prompt is billed as more tokens than on older Claude models. Compare the cost of a whole job, not the price per token.

On 3 June OpenAI announced it is deprecating Agent Builder, its Evals platform and reusable prompt objects. If you built on any of them, plan the move now. Deprecated means a shutdown date will follow.

The 4 Gemini 2.0 Flash and Flash-Lite models stopped on 1 June, as Google announced in February. Calls to them now fail.

From 1 June, OpenAI models including GPT-5.4 and GPT-5.5 can be called through Amazon Bedrock. If your company already works in AWS, it can reach them there.

xAI's coding agent, Grok Build, entered beta in May, and the API gained a compaction endpoint that shortens long conversations. xAI's notes give the month only.

Since 29 May, OpenAI keeps a cached prompt for 24 hours by default, unless your organisation has zero data retention. A prompt you repeat during the day is more likely to get the cache discount.

Claude Opus 4.8 came out on 28 May with a 1 million token context windowContext WindowEverything a model can see at once while it writes an answer, counted in tokens. When the window is full, something has to be left out. by default and effort set to high unless you change it. Sending a non-default TemperatureTemperatureA setting that controls how much chance goes into each word a model picks. Low gives focused, predictable answers; high gives more varied ones. or other sampling setting now returns an error. If your code sets temperature for Claude, take it out before you switch.

gemini-3.5-flash became generally available on 19 May and now answers gemini-flash-latest. On the same day Google opened its own Managed Agents in preview.

OpenAI removed dall-e-2 and dall-e-3 from the API on 12 May. Pictures now come from the GPT image models.

gemini-3.1-flash-lite became generally available on 7 May. The preview name stopped on 25 May, so move code off gemini-3.1-flash-lite-preview.

From 6 May, Claude's API can run several agentsAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. on one job, one handing parts of it to the others, in public beta. Webhooks came the same day, so your code hears when a long job is done instead of asking again and again.

Mistral released mistral-medium-3-5 on 28 April with open weights under a Modified MIT licence. You can download it, not only call it.

GPT-5.5 and GPT-5.5 Pro reached the API on 24 April, 7 weeks after GPT-5.4.

DeepSeek V4 came out on 24 April in 2 sizes, V4-Pro and V4-Flash, and works through both OpenAI-style and Anthropic-style requests. The old names deepseek-chat and deepseek-reasoner now run V4-Flash and are set to stop on 24 July. Change them in your code before then.

OpenAI's new image model, gpt-image-2, reached the API on 21 April. If you make or edit pictures through the API, it is the model to test.

Claude Haiku 3 retired on 20 April, as announced in February. Calls to claude-3-haiku-20240307 now fail.

Claude Opus 4.7 came out on 16 April at $5 per million input tokensTokenThe small piece of text a language model reads and writes, often part of a word. Prices, limits and the context window are all counted in tokens. and $25 per million output tokens. It adds a higher effort level, xhigh, for the hardest jobs, and reads images up to 2576 pixels.

Managed Agents entered public beta on 8 April. Anthropic hosts the AgentAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. for you, so you do not run its loop on your own servers. A command-line tool, ant, came out the same day.

Google released 2 Gemma 4 models on 2 April, gemma-4-26b-a4b-it and gemma-4-31b-it. Gemma models are open, so you can download them and run them on your own machine.

From 1 April the Gemini API has 2 more ways to run a request, Flex and Priority, which trade cost against speed. Check the price page before you pick one for a big batch.

Grok 4.20 went live on the xAI API in March, together with Grok 4.20 Multi-agent, which splits a task across several agentsAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done.. xAI's notes give the month only.

In March Databricks renamed its coding assistant Databricks Assistant to Genie Code. If you look for the Assistant in your workspace, it goes by the new name.

gpt-5.4-mini and gpt-5.4-nano followed on 17 March. They are the smaller, cheaper versions of GPT-5.4, for jobs that do not need the full model.

Since 12 March you can set a spend cap per project in AI Studio, and since 16 March per billing account. If you try the API with a card attached, set a cap first.

Mistral released Mistral Small 4, mistral-small-2603, on 16 March, with a 256,000 token context window.

From 13 March, Opus 4.6 and Sonnet 4.6 read up to 1 million tokensTokenThe small piece of text a language model reads and writes, often part of a word. Prices, limits and the context window are all counted in tokens. at the normal price, with no separate rate limit for long requests. One request can also hold 600 images or PDF pages, up from 100. A long document no longer costs extra per token just for being long.

gemini-3-pro-preview stopped on 9 March, and the name now answers with Gemini 3.1 Pro. Code that used it keeps working but gets a different model, so check its answers.

GPT-5.4 and GPT-5.4 Pro reached the API on 5 March, with a 1 million token context windowContext WindowEverything a model can see at once while it writes an answer, counted in tokens. When the window is full, something has to be left out., built-in computer use and a tool search that loads a tool only when the model needs it.

gemini-3.1-flash-lite-preview came out on 3 March, the first Flash-Lite model in the Gemini 3 family. If you run large volumes of simple jobs, it is the Gemini 3 model to try.

In February Databricks renamed its Multi-Agent Supervisor to Supervisor Agent and made it generally available. It hands parts of a question to other agents and tools and puts their answers together.

Google's new image model, gemini-3.1-flash-image, also called Nano Banana 2, launched on 26 February. It makes and edits pictures from a text request.

gpt-5.3-codex reached the Responses API on 24 February, 6 weeks after 5.2-Codex. If you build on the Codex models, there is a newer one to test.

Since 19 February, the Claude API caches the repeated start of your prompts without you marking it, so repeated text costs less. The same day, Claude Sonnet 3.7 and Claude Haiku 3.5 retired on the Claude API. Haiku 3.5 stays on Amazon Bedrock and Google Cloud, and Haiku 3 follows on 20 April.

gemini-3.1-pro-preview came out on 19 February, with a second endpoint tuned for agents that mix shell commands with their own tools. Gemini 3 Pro's preview shuts down on 9 March, so move to 3.1 before then.

Claude Sonnet 4.6 arrived on 17 February. On the same day, code execution became free when the model uses it with web search or web fetch, so an agent can filter what it finds before reading it.

On 10 February the Responses API gained Skills, a hosted shell the model can run commands in, and server-side compaction. These are the parts an AgentAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. needs to work for longer on its own.

Claude Opus 4.6 came out on 5 February, with a 1 million token context windowContext WindowEverything a model can see at once while it writes an answer, counted in tokens. When the window is full, something has to be left out. in beta and a compaction API that shortens long conversations for you. Setting a thinking budget by hand is now deprecated: the model decides how long to think.

From 3 February, gpt-5.2 and gpt-5.2-codex answer about 40% faster. OpenAI says the model is the same, so answers should not change. Only the wait does.

Structured outputs, which make Claude answer in the exact JSONJSONJavaScript Object Notation - a lightweight data interchange format shape you define, came out of beta on the Claude API on 29 January for Sonnet 4.5, Opus 4.5 and Haiku 4.5. The setting moved from output_format to output_config.format, so older code needs a small change.

Google's Computer Use tool now works with the Gemini 3 Pro and Flash previews, from 29 January. An AgentAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. built on them can look at a screen and click and type its way through a task.

Mistral's coding model Devstral 2.0 moved to paid API access on 27 January. If a coding tool of yours ran on the free tier, check its bill.

Since 21 January, gemini-pro-latest and gemini-flash-latest answer with the Gemini 3 previews. If your code uses a "latest" name, its answers now come from a newer model. Test it, or pin a fixed version.

gpt-5.2-codex, the coding model behind Codex, arrived in the Responses API on 14 January. Developers can now build their own coding tools on the same model.

The Gemini API now accepts files up to 100 MB, up from 20 MB, and can read them straight from Cloud Storage or a pre-signed link. Long recordings and large PDFs no longer need to be cut up first.

Claude Opus 3 stopped answering on the Claude API on 5 January. Calls to claude-3-opus-20240229 now fail. If an old script or tool still names it, move it to a current model.