A new version does nothing to your feature until you point at it. The work is deciding whether to move, and the only honest way to decide is to run your own dataset against both.
A provider shipped a new model or a new dated snapshot. This is an option, not an event, unless an alias you ride now resolves to it.
Meta made muse-voice-transcribe-1.0 available on the Meta Model API for streaming speech to text with speaker diarization. Meta lists the price as $3.00 per 1,000 minutes and offers a WebSocket streaming endpoint and an HTTP batch endpoint.
Summarisation: A transcription step feeds the summariser its input, so the text being summarised changes even though the summarising model did not. Galtea warns against updating ground truth to match model behaviour when the outputs shift.
Extraction: If fields are pulled out of a transcript, a new transcription model changes the input to every fixture. Braintrust tracks every change to a dataset so experiments can pin to specific versions, which is what keeps a before and after run comparable.
Rerun the suiteIf a transcription step feeds a downstream feature, run your existing audio fixtures through this model and compare the transcripts before switching any pipeline.
OpenAI released GPT-6 Astra, which it describes as its most capable model for end to end work, covering reasoning, coding, computer use and tools. It is available in the Responses API and the Chat Completions API.
Agent: A different model plans and calls tools differently. Future AGI describes scoring an agent at four resolutions, end to end, trajectory, step and turn, and writes that a high end to end score can hide bad tool calls, while strong per step scores can still miss that the agent pursued the wrong goal.
Chat assistant: Running the assistant on the new model beside the one you serve is pairwise comparison, one of the four evaluator types LangSmith documents, alongside human review, rule based code functions and LLM as judge.
Summarisation: If a model from this family also runs as your judge, the scoring side is checked against human labels before its verdicts carry weight. Arize writes that a 75 to 90 percent match between judge labels and human labels indicates strong alignment and readiness to scale.
Rerun the suiteAdd gpt-6-astra as a second model in your next scheduled suite run alongside the one you serve today, and read the differences before changing any default.
Google released gemini-3.8-flash as generally available. The entry states it brings improvements for long horizon software engineering.
Agent: Long horizon work is mostly tool calls. Arize Phoenix reconstructs two signals from a trace, tool_path, the ordered tool calls with their arguments, and tool_io, what each tool was called with and what it returned, and warns that a correct looking answer can come from a broken path.
Generation: Generated text has no single correct answer to match against, so the tempting fix when a new model writes differently is to move the expected outputs. Galtea warns against updating ground truth to match model behaviour.
Rerun the suiteQueue gemini-3.8-flash into the next comparison run against the Flash model you use now, and record the result before moving traffic.
Anthropic launched claude-fable-5-1 and claude-mythos-5-1 with a 1M token context window by default, 128k maximum output tokens and always on adaptive thinking, priced at $10 and $50 per million tokens. The notes state that tool_choice types any and tool are not supported and return a 400 error, and that thinking blocks are preserved only for the same model or a newer one.
Agent: tool_choice types any and tool now return an error, so the way the agent is made to call a tool changes. Arize Phoenix reads tool_path, the ordered tool calls with their arguments, and tool_io, what each tool was called with and what it returned, and warns that a correct looking answer can come from a broken path.
Extraction: Forcing a tool call is one way to fix an extractor's output shape, and that request now fails. DeepEval defines a golden as an input plus the expected result, which becomes a full test case once the application produces the actual output, so the same set reruns once the shape is settled.
Revisit the promptSearch your codebase for tool_choice values of any or tool, since those requests now fail, and check any code that replays thinking blocks across models.
Meta made muse-image-1.0 available to developers on the Meta Model API at $0.01 per image. Meta describes anchored composition for consistency across a series of images and reasoning driven edits.
Generation: There is no single correct output to match a generation prompt against, so a fixed prompt set and the recorded outputs carry the comparison. Galtea warns against updating ground truth to match model behaviour.
Rerun the suiteIf image generation sits inside a product flow, run your standing prompt set against this model and keep the outputs as a baseline for the next version.
Google released gemini-3.7-flash. The entry states it brings substantial improvements across software engineering.
Chat assistant: Putting the new Flash model beside the one you pin is pairwise comparison of two application versions, one of the four evaluator types LangSmith documents, alongside human review, rule based code functions and LLM as judge.
Rerun the suiteAdd gemini-3.7-flash to the comparison run you already schedule for Flash tier models, and note where its answers differ from the version you pin.
Meta released Muse-Glimmer-30B, a 30 billion parameter open weight model for local agent workloads, under the Apache 2.0 licence with a 128K token default context window. Meta lists quantized builds and support for vLLM, llama.cpp, ExecuTorch, Ollama and LM Studio.
Agent: An open weight model is pinned by you, so the swap happens when you upgrade rather than when the provider ships. Future AGI writes that agent traces in contemporary frameworks are routinely 15-40 spans deep, so the comparison reads the trace and not the final answer alone.
Rerun the suiteIf you self host a Llama 4 model, run your current suite against Muse-Glimmer-30B on the same hardware and record both quality and throughput before planning a move.
Meta released muse-spark-1.2 with a 1M token context window, priced at $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens. Meta states the model was co-trained with its harness and trained on whole repository generation and other long running tasks.
Agent: Whole repository work and other long running tasks are scored end to end by default. Arize Phoenix writes that end to end evaluation only inspects the two endpoints, the input and the final output, and that reasoning, tool selection and the decision path are where the interesting failures sit.
Generation: Running 1.2 and 1.1 in the same pass is pairwise comparison of two application versions, one of the four evaluator types LangSmith documents.
Rerun the suiteRun your coding and long context cases against muse-spark-1.2 and muse-spark-1.1 in the same pass, and note that the contributor tier is rate limited by tokens in a rolling window rather than by request count.
Anthropic launched claude-opus-5 at $5 and $25 per million tokens, with a 1M token context window as both the default and the maximum and 128k maximum output tokens. The notes state that thinking is on by default, that disabling thinking is only allowed at effort high or below, which they call a breaking change from Claude Opus 4.8, and that Fast mode was removed for Claude Opus 4.7 so those requests return an error.
Agent: Thinking on by default, and a narrower way to turn it off, change how much the agent works out before it acts. Future AGI describes scoring an agent at four resolutions, end to end, trajectory, step and turn, and writes that a high end to end score can hide bad tool calls, while strong per step scores can still miss that the agent pursued the wrong goal.
Chat assistant: Future AGI recommends refreshing a golden dataset immediately after a model update, monthly by default and weekly during rapid change, sampling 200-500 recent production traces stratified by intent and persona and versioning the result as a new artifact.
Generation: If a model from this family also scores your outputs, the judging side moves too. Future AGI writes that a judge model can move 3-5 points on a fixed rubric in a single month without any change on your side.
Revisit the promptFind any request that disables thinking at effort xhigh or max, since that combination is now rejected, and remove the fast mode flag from Claude Opus 4.7 calls.
Google released gemini-3.6-flash and gemini-3.5-flash-lite as generally available. The same entry deprecates the sampling parameters temperature, top_p and top_k.
Classification: Classifiers usually pin sampling parameters in the request builder, so a deprecation here is a change to caller code rather than to the model you chose. Anthropic made the same kind of change from Claude Opus 4.7 onward, where a non default temperature, top_p or top_k returns a 400 error.
Extraction: Removing sampling parameters changes the request every fixture runs through. Braintrust tracks every change to a dataset so experiments can pin to specific versions, which is what keeps the run before the edit comparable with the run after it.
Revisit the promptSearch your request builders for temperature, top_p and top_k on Gemini calls and plan their removal, then rerun the affected cases with the parameters dropped.
Meta announced Muse Spark 1.1, a multimodal reasoning model built for agentic tasks, with what it calls major gains in tool and computer use, coding and multimodal understanding. Meta states it has a 1 million token context window and launched a public preview of the Meta Model API for developer access.
Agent: Changes in tool and computer use land in the middle of the trace. Arize Phoenix reconstructs tool_path, the ordered tool calls with their arguments, and tool_io, what each tool was called with and what it returned, and warns that a correct looking answer can come from a broken path.
Rerun the suiteIf you track Meta models, open an account on the Meta Model API and run your existing suite once so later versions have a baseline to be compared against.
Anthropic launched claude-sonnet-5 at an introductory $2 and $10 per million tokens with a 1M token context window and 128k maximum output tokens. The notes state that adaptive thinking is on by default, that manual extended thinking was removed and returns a 400 error, that non default sampling parameters return a 400 error, and that a new tokenizer produces about 30 percent more tokens.
Classification: Non default sampling parameters now return a 400 error, which breaks caller code without any model swap on your side. Anthropic first made this parameter deprecation from Claude Opus 4.7 onward.
RAG: A tokenizer that produces about 30 percent more tokens changes how much retrieved text fits in one prompt. Ragas defines faithfulness as the share of claims in a response that can be inferred from the retrieved context, so a smaller retrieved context can move the score by itself.
Summarisation: The same input now costs more tokens, so what reaches the summariser can change before the model does anything. Galtea warns against updating ground truth to match model behaviour when the summaries come back different.
Revisit the promptRecalculate token budgets and cost estimates against the new tokenizer, and strip manual thinking and non default sampling parameters from requests before they start failing.
Anthropic launched claude-opus-4-8 with a 1M token context window by default on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. The notes state that the effort parameter defaults to high, that the minimum cacheable prompt length was reduced to 1,024 tokens, that adaptive thinking triggers reasoning only when needed, and that mid conversation system messages are available without a beta header.
Agent: An effort parameter that defaults to high changes how much happens between the request and the answer. Arize Phoenix writes that end to end evaluation only inspects the two endpoints, the input and the final output, and that reasoning, tool selection and the decision path are where the interesting failures sit.
Chat assistant: Mid conversation system messages change the shape of a multi turn request. LangSmith documents four evaluator types for comparing versions, human review, rule based code functions, LLM as judge in reference free and reference based variants, and pairwise comparison.
Rerun the suiteRun your suite once at the default effort of high before you override it, and re-check cache hit rates now that shorter prompts qualify for caching.
Google released gemini-3.5-flash as generally available. The same entry launches Managed Agents in public preview and an antigravity-preview-05-2026 general purpose managed agent.
Chat assistant: Moving a preview pin onto a generally available id is the direction Google's own documentation points, saying most production apps should use a specific stable model and hedging that stable models usually do not change.
Agent: A managed agent in public preview carries a shorter clock than a stable model. Google says Gemini preview models may be used in production but will be deprecated with at least 2 weeks' notice.
Rerun the suiteMove any preview Flash pin onto the generally available id and run the suite once to see whether the outputs shifted.
OpenAI released GPT-5.5 as a frontier model with a 1M token context window, structured outputs and extended prompt caching, alongside gpt-5.5-pro.
Extraction: Structured outputs decide whether extracted fields validate against your schema. DeepEval defines a golden as an input plus the expected result, which becomes a full test case once the application produces the actual output, so the same goldens rerun against the new model.
RAG: A larger window invites more retrieved text into the prompt. Ragas defines faithfulness as the share of claims in a response that can be inferred from the retrieved context, and it scores grounding in that context rather than factual truth.
Rerun the suiteRun your structured output cases against gpt-5.5 and check that responses still validate against your schemas before changing the model you serve.
Anthropic launched claude-opus-4-7 at $5 and $25 per million tokens, the same price as Claude Opus 4.6. The notes state that task budgets launched in beta, that high resolution image input is supported up to 2576 pixels on the long edge, and that an xhigh effort level was added.
Agent: Task budgets and a new effort level change how much the agent does before it answers. Future AGI writes that agent traces in contemporary frameworks are routinely 15-40 spans deep, so the comparison reads the trace rather than the final answer alone.
Extraction: Higher resolution image input means page and screenshot fixtures may be re-encoded, which changes the inputs themselves and not only the model. Braintrust tracks every change to a dataset so experiments can pin to specific versions.
Rerun the suiteRun the suite against claude-opus-4-7 at your current effort level first, then once at xhigh, and record what the extra effort changes before you make it the default.
OpenAI released GPT-5.4 mini and GPT-5.4 nano to the Chat Completions and Responses APIs.
Classification: Weights & Biases documents an evaluation as a dataset of test examples plus one or more scoring functions, used to measure performance against repeatable test cases, compare changes over time and identify regressions. That is the structure a move to a smaller id is judged on.
Rerun the suiteIf you are considering a cheaper id for routing or tagging, run the same labelled set through gpt-5.4-mini and gpt-5.4-nano and compare it against the model you use today.
OpenAI released gpt-5.3-codex to the Responses API. The same entry added the phase parameter, which OpenAI says labels an assistant message as intermediate commentary or final answer.
Agent: Future AGI describes scoring an agent at four resolutions, end to end, trajectory, step and turn, writes that a high end to end score can hide bad tool calls, and warns that strong per step scores can still miss that the agent pursued the wrong goal. A label that separates intermediate commentary from the final answer is read at those resolutions.
Rerun the suiteRun your agent cases against gpt-5.3-codex, and decide whether your harness should record the phase label so intermediate messages stop being scored as answers.
Google released gemini-3.1-pro-preview and launched a gemini-3.1-pro-preview-customtools variant.
Agent: A custom tools variant changes how the agent reaches its answer, not only what it answers. Future AGI describes scoring an agent at four resolutions, end to end, trajectory, step and turn, and writes that a high end to end score can hide bad tool calls, while strong per step scores can still miss that the agent pursued the wrong goal.
Rerun the suiteRun the Pro tier comparison against gemini-3.1-pro-preview, and keep the custom tools variant in a separate run so tool behaviour differences stay visible.
Anthropic launched claude-sonnet-4-6 at $2 and $10 per million tokens with a 1M token context window in beta and extended thinking support. The notes state that web search and web fetch are free when used with code execution, and that web search, web fetch, programmatic tool calling, code execution, tool search, tool use examples and the memory tool are available without a beta header.
Agent: Tools that no longer need a beta header are tools a request can now reach. Arize Phoenix reconstructs tool_path, the ordered tool calls with their arguments, and tool_io, what each tool was called with and what it returned, and warns that a correct looking answer can come from a broken path.
RAG: Where web search or web fetch supplies the documents, the retrieved context changes with the tool path. Ragas defines faithfulness as the share of claims in a response that can be inferred from the retrieved context, and it scores grounding in that context, not factual truth.
Rerun the suiteRemove the beta headers you no longer need from tool requests and rerun the tool using cases, since header changes can alter which tool path a request takes.
Anthropic launched claude-opus-4-6 at $5 and $25 per million tokens. The notes state that adaptive thinking is recommended and manual thinking is deprecated, that an effort parameter with levels low, medium, high, xhigh and max launched, that a 1M token context window is available in beta with long context pricing above 200k input tokens, and that the media limit rose from 100 to 600 images or PDF pages with the 1M context.
Agent: An effort level replaces a thinking budget you used to set by hand, which changes what happens between the request and the answer. Arize Phoenix writes that end to end evaluation only inspects the two endpoints, the input and the final output, and that reasoning, tool selection and the decision path are where the interesting failures sit.
Generation: Moving from a manual thinking budget to adaptive thinking plus an effort level is a prompt change, and the text comes back different. Galtea warns against updating ground truth to match model behaviour.
Revisit the promptSwitch requests from manual thinking budgets to adaptive thinking plus an effort level, then rerun the cases where you previously tuned the thinking budget by hand.
OpenAI released gpt-5.2-codex to the Responses API and describes it as optimized for agentic coding tasks in Codex.
Agent: Arize Phoenix describes reconstructing tool_path, the ordered tool calls with their arguments, and tool_io, what each tool was called with and what it returned, and warns that a correct looking answer can come from a broken path. Those are the signals a coding agent is read on.
Rerun the suiteIf a coding or agent feature sits on a pinned Codex id, run your trajectory cases against gpt-5.2-codex before you point anything at it.
Google launched gemini-3-flash-preview. The entry describes frontier class performance and upgraded reasoning capabilities.
Summarisation: A summariser pinned to a preview id runs on a shorter clock. Google says Gemini preview models may be used in production but will be deprecated with at least 2 weeks' notice.
Rerun the suiteAdd gemini-3-flash-preview to your Flash tier comparison run, and keep the preview id out of production pins until a stable id ships.
OpenAI released GPT-5.2 as its flagship model, stating improvements in reasoning, vision, code generation and spreadsheet understanding, alongside gpt-5.2-chat-latest. The same day OpenAI released client side compaction for long running conversations.
Extraction: Vision and spreadsheet inputs are where extracted fields come from. DeepEval defines a golden as an input plus the expected result, which becomes a full test case once the application produces the actual output, so the document fixtures rerun unchanged against the new model.
Generation: Running the new flagship beside the snapshot you pin is pairwise comparison of two application versions, one of the four evaluator types LangSmith documents.
Rerun the suiteRun your suite against gpt-5.2 alongside the GPT-5.1 snapshot you pin, and pay attention to the vision and document cases the entry calls out.
Anthropic launched Claude Opus 4.5, which the release notes describe as its most intelligent model, combining maximum capability with practical performance. The release notes name the family alias claude-opus-4-5. Anthropic's model deprecations table lists the dated snapshot claude-opus-4-5-20251101 as Active, with retirement not sooner than 24 November 2026.
Rerun the suiteA new model is an option rather than an event, unless an alias you already call now resolves to it. Run your own dataset against the model you use today and against this one before deciding to move.
Google launched gemini-3-pro-preview. The entry describes state of the art reasoning and multimodal capabilities.
Chat assistant: A first baseline on a preview id is a baseline with a short life. Google says Gemini preview models may be used in production but will be deprecated with at least 2 weeks' notice.
Rerun the suiteSet up a first baseline run against gemini-3-pro-preview so later Gemini 3 versions have something to be compared with.
OpenAI released GPT-5.1 as a flagship model optimized for steerability, code generation and agentic workflows, and released gpt-5.1-codex and gpt-5.1-codex-mini to the Responses API. The same day OpenAI launched extended prompt cache retention of up to 24 hours.
Agent: Agentic workflows are judged across the whole run. Future AGI writes that agent traces in contemporary frameworks are routinely 15-40 spans deep, naming the OpenAI Agents SDK, LangGraph, CrewAI, AutoGen and Google ADK.
Chat assistant: A family alias resolves to whatever the provider points it at. Google documents this for Gemini, where a latest alias points at the most recent release of a model variation and is hot swapped with every new release, which is the reason to pin an explicit id.
Re-pin the versionPin the exact GPT-5.1 id you intend to serve rather than a family alias, and rerun the prompts you had tuned against GPT-5 before switching.
Anthropic launched Claude Haiku 4.5, which the release notes describe as its fastest and most intelligent Haiku model, with near-frontier performance. The release notes name the family alias claude-haiku-4-5. Anthropic's model deprecations table lists the dated snapshot claude-haiku-4-5-20251001 as Active, with retirement not sooner than 15 October 2026.
Rerun the suiteIf you are considering this model to cut cost or latency, the comparison that matters is your own dataset scored on both models, not a published summary of either.
OpenAI released GPT-5 Pro, which it describes as a version of GPT-5 that uses more compute to think harder. The same DevDay entry launched the v1/videos API with the Sora 2 and Sora 2 Pro models, and released GPT-Realtime mini, gpt-audio-mini and gpt-image-1-mini for cost efficient speech and image work. It also launched Agent Builder, ChatKit and Trace Evals.
Generation: A release that adds generation models gives a team a new option rather than a forced change. Galtea warns against updating ground truth to match model behaviour, so the outputs you already recorded stay the comparison point when a new id is tried.
Rerun the suiteNothing moves until you adopt one of these ids. If you plan to try gpt-5-pro, run your standing prompt set against it first and keep the current model's outputs beside it.
Anthropic launched claude-sonnet-4-5-20250929, which it describes as its best model for complex agents and coding. The notes state that a new stop reason model_context_window_exceeded was added, that the memory tool launched in beta and that context editing launched in beta.
RAG: Context editing and a new stop reason change what actually reaches the model on a long input. Ragas defines faithfulness as the share of claims in a response that can be inferred from the retrieved context, so the score follows whatever survived.
Agent: A memory tool adds calls that the final answer never shows. Arize Phoenix reconstructs tool_path, the ordered tool calls with their arguments, and tool_io, what each tool was called with and what it returned.
Rerun the suiteAdd handling for the model_context_window_exceeded stop reason so long inputs fail visibly, and rerun the long context cases that previously truncated.
Change-log entries restate provider announcements with sources and dates. Impact notes cite public practitioner reports. The site publishes no measurements of its own on any model.
Questions about new versions
How is a change classified as a new version?
From what the provider wrote, not from a guess about severity. Where a provider states that weights are unchanged, the entry is not classified as a weight change, and where an unchanged model string starts returning something else, it is.
Is this list complete?
It covers the model families we track for customers, inside a twelve-month window, where the announcement is reachable on the provider's own domain. Anything we cannot reach there is not published, however widely it was reported. Send us a provider link and we will add it.
Do you say what this did to real features?
No. The impact notes say what this kind of change typically touches, each citing a published practitioner source. What it did to your feature is a question only a rerun against your own dataset answers.