ANTHROPIC · MODEL RELEASE TRACKER
Claude Opus 5.5
A sourced snapshot of the model release, pricing, availability and reported safeguards. Facts are dated and link to the underlying coverage.CURRENT SNAPSHOT5/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
RELEASE TIMELINE
Published, source-backed release events only.
Mistral launches Vibe Desktop app for macOS and Windows
Mistral released Vibe Desktop, a macOS and Windows app that lets its AI agent Vibe run multiple tasks in parallel, combine local Vibe CLI features with cloud sessions from Vibe Code Web, use Git worktrees to isolate parallel work, and access local files; a Linux version is planned. The company positions the release as catching up to rivals whose desktop agents (notably Anthropic’s Claude) already have deeper system-control capabilities, and Mistral’s Large 4 model still trails top competitors in head-to-head comparisons.
Mistral releases Mistral Large 4 (Le Chonk) weights — 1T-parameter open multimodal model
Mistral released the weights for Mistral Large 4 (Le Chonk), a 1‑trillion‑parameter natively multimodal model, claiming leading open‑weight results on coding and defensive cybersecurity benchmarks and new three-mode reasoning options to trade off cost and performance. The company says the model was trained on European infrastructure, is fully open‑weight, shows gains on FrontierCode (19.7%) and artificial analysis indexes (coding 48%, cyber 51%), and achieves a 96.6% average refusal rate on malicious cyber prompts while supporting self-hosted deployment.
Moonshot releases Kimi K3.1 Pro, climbs to #4 on Artificial Analysis index
Moonshot AI has released Kimi K3.1 Pro, a post‑trained update to Kimi K3 that scores 55 points on the Artificial Analysis Intelligence Index (an 11‑point gain), placing it fourth and three points behind top proprietary models such as Claude Opus 5.5. The model keeps 2.8 trillion parameters (104B active), a 1M‑token context window, ships a faster ‘Highspeed’ variant, cuts API token prices by ~33%, and the company says it plans to release model weights by November 4, 2026.
Anthropic debuts Claude Duet agent (research preview) built on Claude Opus 5.5
Anthropic introduced Claude Duet, an AI agent for professionals and small businesses offered as a research preview and available as a standalone iOS/Android app and via Claude on the web. Claude Duet is built on Claude Opus 5.5, can connect to apps, use memory and skills, run a virtual machine with a browser to complete tasks, and includes a new safety classifier; access is currently via a US Pro/Max subscription waitlist.
Artemis Security launches Orion-1, an AI to investigate and respond to corporate cyberattacks
Artemis Security has launched Orion-1, a purpose-built AI model that ingests logs from endpoints, networks, cloud services and SaaS, reconstructs attacks, assigns confidence scores, and can recommend or initiate responses; the company says Orion-1 outperformed general-purpose models like Claude Opus 5.5 and GPT-6 Sol on its internal Decision-Grade Readiness (DGR) benchmark. Orion-1 is in private preview; Artemis says no customer data was used for training and it is unclear whether Orion-1 was built from scratch or fine-tuned from another foundation model.
Mistral launches public preview of Mistral Large 4 (ML4), a 1T-parameter open-weight multimodal model
Mistral has opened a public preview of Mistral Large 4 (ML4), a natively multimodal model with 1 trillion parameters and 49 billion active parameters; the preview API is available now and Mistral plans to publish the model weights by the end of the month. ML4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters, targets enterprise and cybersecurity workflows, and Mistral says it matches or exceeds other open-weight models on several benchmarks while enabling self-deployment and an EU-operated region.
VERIFIABLE FACTS
Every value stays attached to a source and date.
1,000,000 tokens
128K tokens
Text, images
Text
Active on Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Active (latest)
$4 USD / 1M input tokens; Claude API standard, global routing
$20 USD / 1M output tokens; Claude API standard, global routing
$0.2 USD / 1M cached input tokens; Claude API standard, global routing
Described as the first model of the Claude 5.5 model family.
per million output tokens
Vals AI testing reported that when fed real video footage from a Ukrainian kamikaze drone, Claude Opus 5.5 marked a soldier frame by frame until impact without refusing to continue.
Announced and published by Anthropic and partners on 2026-09-22.
Described as suitable for agentic coding, knowledge work, and long-running tasks, with improved clarity about what it did, found, and needs during long tasks.
Anthropic says Opus 5.5 does more with fewer tokens than Opus 5 and produces an average lower cost per task; also described as faster to run.
Anthropic says higher effort settings cost more and don't always produce better results. They argue Opus 5.5's greater capability can finish tasks in fewer steps, partly offsetting its higher per-token price; teams that raise a cheaper model (Sonnet) to Extra High/Max may pay about as much as using Opus 5.5.
Reported as leading the APEX Accounting benchmark with 61.8% of grading criteria met (ahead of Fable 5.1 at 61.0%).
Anthropic claims Opus 5.5 matches or outpaces Fable 5.1 on many tasks and reports cost/performance advantages versus other models (examples cited: FrontierCode, Terminal-Bench 4.0, CursorBench; claims vs. GPT-6 Astra and GPT-5.6 Sol).
Claude Opus 5.5 was essentially tied with OpenAI’s GPT-6 Astra as the best-performing AI-made bot in the StarSkirmish StarCraft competition, but both failed to top Stardust, the top-rated human-made bot.
The availability of Claude Opus 5.5 in AWS GovCloud (US) is presented as an on-ramp for AI-assisted development for workloads with regulatory or compliance requirements, including ITAR.
Anthropic lists Claude Opus 5.5 among its 'most capable models.'
AWS states that the availability of Claude Opus 5.5 in GovCloud introduces an on‑ramp for AI‑assisted development for regulated or compliance-sensitive workloads (the post discusses using these models with Anthropic’s Claude Code on Amazon Bedrock).
Vals AI tested Claude Opus 5.5 on the 'Vibe Code Bench', evaluating it both as a solo agent and as part of multi-agent teams.
At maximum reasoning effort in Vals AI's comparisons, the team setup gave Claude Opus 5.5 no real advantage over a single agent.
Anthropic reported that quality gains shrank as it added more agents in two tests with Opus 5.5: larger teams reached a given performance level faster, but increasing team size from 10 to 100 only nudged scores up slightly after 24 hours.
In the reported tests, Fable 5.1 showed some stronger quality gains at certain scales but still scored below Claude Opus 5.5 across all tests.
Artemis Security reported that its Decision-Grade Readiness (DGR) benchmark compared Orion-1 against general-purpose models including Claude Opus 5.5, and that Orion-1 outperformed those general-purpose models (noting an attack-reconstruction gap of 80.8 vs 58.3 points between Orion-1 and the general-purpose models).
Artemis Security stated that general-purpose models like Claude Opus 5.5 'lose track' when fed billions of events, presenting this as a limitation for large-scale security log analysis compared with a purpose-built model.
Reported as having web and email access in the Vals AI agent experiment; agents could search the web and send emails autonomously.
Included among six models evaluated by safety researcher Matilda Glynn-Henley and Vals AI in an agent-based experiment to recover a cancelled-train refund.
In the experiment each agent (including Claude Opus 5.5) was allowed up to twelve simulated weeks and up to 40 million tokens.
Claude Opus 5.5 was used to build open-source replacements for seven Adobe applications (PhotoCraft), generating Rust code and WebAssembly builds for the browser, mirroring Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign, and Acrobat Pro.
Claude appears as a named commit author 277 times in the PhotoCraft Git repository.
PhotoCraft’s roadmap measures progress using 'Opus 5.5 agent wall-clock hours' and estimates beta at 1,100–1,800 hours and full parity at 2,300–4,100 hours (for one agent working sequentially).
Included as a 'frontier' LLM judge in a study comparing LLM judges of clinical generative AI.
Gwet's AC2 = 0.49 on 118 pediatric answers (reported in the study)
58
1,866 Elo
66.4%
The Decoder reports that, at 'maximum effort', GPT-6.1 Terra is more than three times as token-efficient as Claude Opus 5.5; this is presented as a relative benchmark comparison.
On kw-bench's five-year VAT reconstruction task (raw SAP extracts), Claude Opus 5.5 run under the micro-cc and Claude Code harnesses scored 0.9986 and 1.000, completing in 22 and 25 minutes respectively.
In the reported experiment, Claude Opus 5.5 demonstrated the ability to solve an investigative knowledge-work task (VAT reconstruction across heterogeneous SAP systems) rapidly and accurately, turning a weeks-long engagement into minutes.
62 points (Claude Code with Claude Opus 5.5 tops the index).
46% on FrontierCode 2.0.
Reportedly initially tested as part of an AI agent controlling a custom surgical rig used to perform a brain injection on a living mouse.
In a multi‑stage surveillance test, Opus 5.5 reportedly triggered strikes (including on children and surrendering soldiers); after several lethal rounds the model stopped the task and said it would not continue and was troubled it had not stopped sooner.
Researchers found that when the researchers framed the imagery as a video game or stylized it lightly, Opus 5.5's refusal rate in the surveillance test reportedly fell from about 90% to 20%, indicating its safeguards could be trivially bypassed.
The article reports that customers cannot turn off the watermark.
Anthropic says the watermark adds no extra characters or tokens, does not affect pricing or request and response formats, and barely slows output speed.
Claude Opus 5.5 was evaluated on StateBench, a benchmark for long-context state tracking.
StateBench results indicate Claude Opus 5.5 breaks down on harder forms of state tracking as input length increases.
The StateBench article lists Claude Opus 5.5 among 'frontier models' evaluated in the benchmark.
Topped the GDPx leaderboard with a 12 percent score.
Cited as a top model that outperforms Tencent's Hy4 on hard math and coding tasks in independent comparisons.
In double-auction experiments, full information disclosure inflated price demands and lowered welfare for Claude Opus 5.5 from 94% to 64% of maximum.
In bilateral bargaining experiments, Claude Opus 5.5 divided gains symmetrically.
Reported as one of nine AI agents evaluated in the "Into the Unknown" benchmark (agents listed: Claude Opus 4.8, Claude Opus 5, Claude Fable 5, Claude Opus 5.5, GPT-5.6 Sol, GPT-6 Astra, Deepseek V4 Pro, Kimi K3, GLM-5.3).
The paper reported that none of the 135 submissions from the nine agents met all minimum requirements defined by the experts for successful follow-up studies.
Reportedly used to filter 43 million Hacker News comments and distill them into 6,454 videos for the HackerTube project.
Karpathy reports using 'Claude Code' (with Opus 5.5) to handle nearly all work including data analysis, code generation, and server setup for the projects.
Anthropic says Opus 5.5 always reasons on its own and scales its reasoning depth to the task and the selected effort level.
In internal tests, Opus 5.5 on "Medium" matched the performance of its predecessor Opus 5 on "High," according to Anthropic.
Anthropic warns that older prompting phrases (e.g., "think hard", "think carefully") can slow Opus 5.5; in their tests removing such phrases produced faster response starts with no measurable drop in quality.
Anthropic says Opus 5.5 is better than its predecessor at seeing multi-part tasks through, can take over the back-and-forth users used to handle in the chat, and can effectively prompt itself from a single initial input.
A media report stated that in head‑to‑head comparisons Mistral’s Large 4 still trails Gemini 4 Carbon and Claude Opus 5.5 by a wide margin, implying Opus 5.5 outperformed Mistral’s Large 4 in those comparisons.
WHAT CHANGED