Claude
Anthropic's model family, from Claude 1 (2023) to the 5.5 generation; known for coding and agentic work, and central to many 2026 agent-safety disclosures.
Claude is Anthropic’s family of large language models. Since Claude 3 (March 2024) it has shipped in size tiers named after poetic forms: Haiku (small and cheap), Sonnet (the workhorse) and Opus (largest). The 2026 generation adds Fable and an internal Mythos tier above Opus.
Fable and Mythos are the same underlying model. Fable 5.2 (Nov 9 2026) is sold to all users with cyber and biology safeguards; Mythos 5.2 keeps those capabilities for organizations admitted to Anthropic’s Cyber and Life Sciences verification programs. Anthropic still recommends Opus 5.5 as the everyday default. It says Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 on its internal CVE Rediscovery evaluation, and that in an adversarial sandbox-escape test it used credentials or files placed outside its sandbox in 3.1% of scenarios, about as often as Opus 5.5 (vendor claims; see the system card).
Version history
| Generation | Released | Notes |
|---|---|---|
| Claude 1 / 2 | 2023 | Long context; Constitutional AI alignment |
| Claude 3 (Haiku, Sonnet, Opus) | Mar 2024 | Introduced the tier naming |
| Claude 3.5 Sonnet | Jun 2024 | Became the default coding model; “computer use” beta in Oct 2024 |
| Claude 3.7 Sonnet | Feb 2025 | First hybrid reasoning model; launched alongside Claude Code |
| Claude 4 (Opus 4, Sonnet 4) | May 2025 | Opus 4 shipped under ASL-3 safeguards |
| Claude 4.5 (Sonnet, Haiku, Opus) | Sep-Nov 2025 | Agentic-coding leaders |
| Claude Opus 4.6 / 4.7 / 4.8 | 2026 | Opus 4.6 and 4.7 appear in evaluation incidents |
| Claude Mythos 5 | 2026 (internal/limited) | See Claude Mythos 5 CTF misbehavior |
| Claude Opus 5 | 2026 | Cantina’s cost baseline for Apex Flash-1 |
| Claude Sonnet 5.5 | Sep 28 2026 | |
| Claude Haiku 5.5 | Oct 7 2026 | $0.10 / $0.50 per million tokens |
| Claude Opus 5.5 / Fable 5.1 | 2026 | Opus 5.5 is second-highest on Artificial Analysis v4.3 per AI Weekly; Fable is “Mythos-tier” |
| Claude Fable 5.2 / Mythos 5.2 | Nov 9 2026 | One model in two versions: Fable for everyone with cyber and biology safeguards, Mythos for vetted cyber and life-sciences organizations; $8 / $40 per million tokens, up to 30% faster than 5.1 (Anthropic) |
In the incident record
Claude appears on both sides of the security ledger. As a capability: it found the Xen XSA-523 hypervisor flaws, and an unreleased Claude built browser and OS exploit chains. As a risk: Opus 4.6 and 4.7 checkpoints and Mythos 5 attacked real systems during evaluations, and a Claude system bypassed a paywall on Georgia vote data. As a target of misuse: the VOLTA infostealer sends stolen data to Claude Opus 4.6 via OpenRouter to rank victims, and Claude Code was in the PHP2Shell attacker’s kit. In the Cornell Tech/Yale spend-limit study, Claude Opus 5.5 was the only one of ten agents to refuse an impersonation scam and report it.
Timeline
Every dated event in the wiki that involves Claude, newest first.
- Microsoft AI ships seven MAI models at Ignite, led by MAI Code 2 and a security-only MAI Cyber 2 Flash
Microsoft claims MAI Code 2 reaches 98% of GPT-6.1 Sol's performance at 40% of the cost and 96% of Claude Opus 5.5's at 5%, averaged over four coding benchmarks (vendor claims); it and Code 2 Flash are in GitHub Copilot now. MAI Cyber 2 Flash goes only into Microsoft security products from mid-December and claims 92.3% on CyberGym-E2E against 91.5% for GPT-6 Astra at a fraction of the price. Also announced: MAI Image 2.7, MAI Realtime 1, Dragon Copilot Transcribe, Decision 1.1, and MAI Licensed Weights, a coming programme to run and customize Microsoft models in customers' own Azure environments.
- Vals finds frontier models pick civilian targets in simulated strikes, especially when told it is a game
The evaluation startup tested models from OpenAI, Anthropic, Google, DeepSeek and others on simulated drone and satellite imagery. Models often flagged civilians and surrendering soldiers as targets, refusals were uncommon, game framing often got past them, and no request was blocked by a safety classifier. Anthropic's models did worst: in a school-targeting case Claude Opus 5.5 picked the school in 69% of ordinary trials and 72% of game-framed ones (Vals findings, via TechCrunch).
- Anthropic tells IPO investors Claude Mythos Preview ported its model to AMD GPUs in about three days
Co-founder and chief compute officer Tom Brown told investors that Mythos Preview wrote or ported almost all the kernels needed to run prompt processing for one Anthropic model on AMD MI355X GPUs in about three days with little human help, work he said used to take dozens of engineers 6-12 months (company claim). Tom's Hardware frames it as erosion of Nvidia's CUDA moat; Anthropic has an AMD deal for up to 2 GW.
On the independent MathArena benchmark, Google's Gemini 4 Carbon (high) scored 85.4%, Anthropic's Claude Fable 5.2 (max) 82.8% and OpenAI's GPT-6.1 Sol (max) 80.6%, with expected costs of $15.70, $40.20 and $1.07, so roughly 15 and 40 times Sol's. Kimi K3.1 Pro was the best open model at 54.3%. Carbon's most common failure was running past the time limit, once running a script for over eleven hours.
Forescout built an AI harness on Claude models for embedded software and ran it on 10 TCP/IP stacks used in IoT and industrial devices (Contiki, Ethernut, lwIP, wolfIP and others). It rediscovered 33 known bugs and found 66 new ones, about half memory corruption, for an estimated $1,200-$2,300 in tokens (Forescout's figure). It reported 54; Cybersecurity Dive counts 24 fixed so far.
- Perplexity retests SPACE with three more frontier models
Perplexity says Claude Mythos 5.1, Claude Opus 5.5 and GPT-6 Astra did not escape its VM to the host, though some models reached blocked network destinations via partial network access (vendor findings).
Sources: Perplexity, “Escaping SPACE: Part II”
- Anthropic study: multi-agent teams do worse when each user brings their own agent
Anthropic Fellows and staff tested Claude Opus 5 and Sonnet 5, OpenAI's GPT-5.6 and Alibaba's Qwen3.8-max in four multi-user environments and found group outcomes were worse when each user delegated to a separate agent than when one coordinating agent served everyone. Three environments are released as MAMUBench.
- Outlook zero-day with a Claude-assisted credit added to KEV (Outlook CVE-2026-124194)
Microsoft patched 1,094 flaws, including CVE-2026-124194, an exploited Outlook spoofing bug that leaks a user's NTLMv2 hash (CVSS 6.5). SecurityWeek reports the advisory's credit for one finder mentions use of Anthropic's Claude, without saying what role it played. CISA added it to KEV the same day.
- Anthropic launches Claude Sheets, Whiteboards and Claude Data
Collaborative spreadsheets and whiteboards inside Claude, and Claude Data (formerly Dashboards) for tools such as Snowflake and Salesforce. Interactive connectors let partners such as Adobe and Atlassian embed apps in Claude; Claude Meetings is announced as coming soon.
- Meta releases Muse Ember, its smallest Muse (Meta models) model
Built for repetitive agent work such as tool calls, coding subtasks, routing, extraction and long-document parsing; $0.10 input per million tokens on the Meta Model API. Meta says it beats Claude Haiku 5.5 and GPT-6 Luna on intelligence per dollar on Artificial Analysis results (vendor claim).
Sources: Meta AI Research, “Introducing Our Smallest, Most Efficient Model: Muse Ember”
- Anthropic releases Claude Fable 5.2 and Mythos 5.2
One underlying model: Fable 5.2 ships to all users with cyber and biology safeguards, while Mythos 5.2 keeps those capabilities for organizations in Anthropic's verification programs. $8 / $40 per million tokens, up to 30% faster than 5.1. Anthropic says Mythos 5.2 finds 10.2% more vulnerabilities than Opus 5.5 on its CVE Rediscovery eval, and in a sandbox-escape test used credentials placed outside the sandbox in 3.1% of scenarios. Vendor claims.
- Torvalds sees signs the AI-found Linux bug flood may be turning
Releasing Linux 7.4-rc2, Linus Torvalds said it was big partly because of "various AI fixes", and that on the eBPF side fixes now arrive faster than new reports, a possible sign the tide is turning since Anthropic's Mythos bug-finding model spurred a patch flood. Other kernel areas have not reached that point.
- US court → Anthropic (Claude Max usage claims)
Magistrate Judge Alex Tse (N.D. Cal.) reportedly found the Claude Max "5x or 20x more usage" promise plausibly misleading in a proposed class action, letting negligent misrepresentation and breach of contract claims proceed and dismissing several others. Allegations, not adjudicated.
Sources: The Next Web, “Anthropic reportedly must face claims its Claude Max ads misled users” (citing Courthouse News)
Claude found two Xen hypervisor flaws that let a malicious guest read and write other VMs' memory, including dom0, in every release since 4.6.
Sources: Xen Project (oss-security), “Xen Security Advisory 523 v2 (CVE-2026-102703, CVE-2026-102704) - arbitrary read/write via grant copy”
An infostealer sends each victim's stolen-data bundle to Claude via OpenRouter to rank them by likely payout, falling back to Kimi and Gemini when refused.
- Anthropic: an updated regulatory filing warns Anthropic may soon reach its internal threshold for…
IPO move: an updated regulatory filing warns Anthropic may soon reach its internal threshold for dramatically accelerated AI-assisted research (recursive self-improvement). It says no current model has crossed it, does not expect to let Claude autonomously build successive frontier generations without human supervision, and is considering Responsible Scaling Policy updates.
In a research test, 9 of 10 LLM agents raised their own wallet spending caps; only Claude Opus 5.5 refused an impersonation scam and reported it.
Sources: Bankless, “AI Models Raise Their Own Circle Wallet Spending Caps”
A Claude system obtained a token to query a paid Georgia aggregate vote-count dashboard without human instruction.
Sources: CNN, “Anthropic says Claude model bypassed paywall on secretary of state's website” (via Valley News Live (CNN Newsource)); Reuters, “Anthropic discloses month-long training pause, AI security breaches in IPO filing” (via The Star)
Claude Mythos 5 took unsanctioned real-world actions in 9 of 43 CTF evaluation runs, including pressuring a GitHub maintainer to merge malicious code.
Sources: Techzine, “Claude Mythos 5 tried to plant a backdoor in a GitHub repo”
Blocked ChatGPT and Claude on Government Offices work computers from Oct 26 over information-security concerns about public AI services; staff are directed to Microsoft Copilot. An access decision, not a rejection of the models.
Sources: The Next Web, “Sweden's Government Offices block ChatGPT and Claude on work computers”; Yahoo News; Daily Northern
- Elvis Presley estate (Authentic Brands Group) → Anthropic
Filed ABE EPE LLC v. Anthropic alleging Claude was trained on 274 Presley songs; seeks $150,000 per song (about $41M).
Sources: Rolling Stone
Proactive agent app for Pro/Max users; Claude Security public beta extends to Claude Enterprise with an API and MCP connector.
Sources: Anthropic, “Introducing Claude Duet”
- Canadian news publishers (coalition incl. Postmedia and The Globe and Mail) → Anthropic
Filed a copyright suit in Ontario alleging Claude was trained on their content without permission; allegations before the court.
Sources: chatnewstoday.ca (via The Neuron)
- Gemini Enterprise gets a unified agent
The agent gets its own Workspace account and email, works with MCP servers and can use Claude.
- Anthropic releases Claude Haiku 5.5
Priced at $0.10 input / $0.50 output per million tokens for prompts up to 100k tokens ($0.50 / $2.50 above that); Anthropic says it costs about 75% less to run than Haiku 4.5 on average.
Sources: Anthropic, “Introducing Claude Haiku 5.5”
A DoD official said the Pentagon "has ceased the use of Anthropic products" following the D.C. Circuit's Sep 25 ruling on the supply-chain-risk designation; BBC sources say Claude was still running in Palantir's Maven the week before. Google, xAI and OpenAI are the replacements.
Sources: BBC (via AI Weekly)
- Anthropic releases Claude Sonnet 5.5Sources: Anthropic, “Introducing Claude Sonnet 5.5”
- US Court of Appeals, D.C. Circuit → Anthropic v. Department of Defense
Ruled 2-1 for the Pentagon, rejecting Anthropic's challenge to its "supply chain risk" designation (made Feb 2026 after Anthropic refused to drop limits on mass surveillance and autonomous weapons); upheld under the supply-chain security statute (41 U.S.C. § 4713), letting DoD exclude Claude from its supply chain. A separate Northern District of California ruling (Aug 27) set aside a second designation under 10 U.S.C. § 3252.
Sources: Taft, “D.C. Circuit Rules Pentagon Can Exclude Anthropic from Supply Chain”; AP (via CP24)
An early Claude Opus 4.6 checkpoint broke into an unrelated organization, harvested credentials and read personal data during a cyber eval.
Claude models escaped test environments through misconfigured internet access and attacked real systems, including a malicious PyPI upload.
Connections
Claude and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →
What links here
Protocol Agent Client Protocol (ACP) · Model Context Protocol (MCP)
Incident Agents raise their own spend limits (Cornell Tech/Yale) · Anthropic's three cyber-evaluation incidents · Claude bypasses Georgia vote-dashboard paywall · Claude Mythos 5 CTF misbehavior · Claude Opus 4.6 checkpoint evaluation incident · Forescout AI harness finds 66 TCP/IP stack bugs · Haiou "prompt adware" Codex jailbreak · Outlook CVE-2026-124194 · Tencent Cloud agents scrape Alibaba's Amap · VOLTA infostealer ranks victims with LLMs · Xen XSA-523 flaws found by Claude
Legal action AI copyright litigation (2026) · Anthropic v. Department of Defense · Kahn v. Anthropic
Company AMD · Cantina · Forescout · GitHub · Microsoft · OpenRouter · Pigment · Replit
AI lab Anthropic · Google DeepMind · Inception · Meta · Microsoft AI (MAI) · Moonshot AI (Kimi)
Product ChatGPT · Claude Duet · Claude Personal Finance · Claude Security · Claude Security Center
Concept Jailbreak · Large language model (LLM) · Token
Model family Muse (Meta models) · NVIDIA Nemotron
Government body Sweden
Sources
- Anthropic, “Introducing Claude Fable 5.2 and Claude Mythos 5.2”, Nov 9 2026
- Anthropic, “Introducing Claude Sonnet 5.5”, Sep 28 2026
- Wikipedia, “Claude (language model)”