What’s the big deal
Eighteen months ago, there was one Claude model and it was good at writing. Today there are three core models that compete at the very top of every major benchmark, plus a Mythos-class tier above them whose first version was withheld entirely and whose second was briefly pulled off the market by a government order. Fable 5.1, the September update to that public Mythos-class model, remains Anthropic’s premium flagship, but Opus 5.5, launched 22 September at 40% of Fable’s price, now performs at its level on most work and leads the Artificial Analysis Intelligence Index, with Sonnet 5.5, a faster 28 September upgrade at half of Opus 5.5’s price, close behind. No single model wins everything: OpenAI’s GPT-6 Astra and Google’s Gemini 4 Argon tie Fable 5.1 on that index, Gemini 4 Argon took first on Arena’s human-preference text board on early votes, and OpenAI’s far cheaper GPT-6.1 Sol matches Claude on coding at equal effort, while Claude leads on knowledge work, writing, and the subjective quality that keeps users coming back. The practical effect is that frontier-level intelligence at the $2 per million token price point would have required the $15 flagship tier a year earlier.
What it is
Claude isn’t a single AI, it’s a family of models, each built for different needs. Think of them as tiers of the same brain: the more powerful the model, the better it handles complex or high-stakes work, but the slower and more expensive it is to run. All can read text and images, work in multiple languages, and handle long documents. What separates them is the depth of reasoning, the speed, and the price.
The current lineup
Opus 5.5
Released September 22, 2026, succeeding July’s Opus 5The most capable model in the everyday lineup, sitting just below the Mythos-class Fable tier at roughly 40% of its price, and performing at the level of Fable 5.1 on most work, per Anthropic, with a real-world gap the company says is narrower than benchmark scores suggest (Anthropic rates it comparable to Mythos 5.1 in cybersecurity and biology). It leads on complex reasoning, scientific questions, and tasks that require sustained focus across long, multi-step work: professional software engineering, multi-hour research, or advanced agents that handle office work autonomously. It is the default backbone of Claude Code, supports Agent Teams and Dynamic Workflows, and adds an adjustable effort setting plus a beta for changing tools mid-conversation. Priced at $4/$20 per million tokens, about 40% cheaper than Opus 5 on typical workloads, roughly 30% faster, with a Fast Mode at 2.5x speed for $8/$40. Choose it when you want near-frontier quality without Fable’s premium price; reach for Fable only when you need the absolute frontier.
Sonnet 5.5
Released September 28, 2026, succeeding June’s Sonnet 5The best combination of speed and intelligence, and the default model for most people on claude.ai, including Free and Pro. Anthropic’s most agentic Sonnet yet: it plans multi-step work, drives browsers and terminals, and runs autonomously at a level that used to need an Opus-class model. Same $2/$10 per million token price as Sonnet 5, but runs roughly 30% faster and costs up to 30% less for most work, with new cyber safeguards and reasoning-extraction defences like those built for Anthropic’s most capable models. 1M token context window. For most day-to-day work this is still the model to reach for; choose Opus 5.5 for the hardest, highest-stakes tasks.
Haiku 5.5
Released October 7, 2026, succeeding 2025's Haiku 4.5The fastest and cheapest model in the lineup, at roughly 75% lower average cost than Haiku 4.5: $0.10/$0.50 per million input/output tokens for prompts up to 100k tokens, rising to $0.50/$2.50 above that. The first Haiku-class model with an adjustable effort setting, and now matches the core Opus and Sonnet tiers on a 1M token context window, 128K max output, and adaptive thinking. Typically used behind the scenes to power high-volume automated workflows, real-time applications, and fast sub-agents inside larger systems.
Mythos class: Fable 5.1 and Mythos 5.1
Released June 9, 2026; suspended June 12, restored July 1; updated September 1A tier above Opus. Fable 5.1 is the public version and Anthropic’s premium flagship for the most demanding reasoning and long-horizon agentic work, though the cheaper Opus 5.5 now matches it on most work and outscores it on the Artificial Analysis Intelligence Index (58 to 53). Mythos 5.1 is the same underlying model with cybersecurity and biology safeguards lifted, restricted to Project Glasswing and Anthropic’s other trusted-access programmes. Both are priced at $10/$50 per million tokens, with cache reads cut to $0.25/MTok in the September update. A US government export-control directive forced the original versions offline worldwide on 12 June; the controls were lifted on 30 June, and Fable 5 returned globally on 1 July after Anthropic trained a new classifier to block the jailbreak that triggered the ban. See the Fable and Mythos topic page.
Specifications
Context windows: Opus 5.5, Sonnet 5.5, and Haiku 5.5 all support 1M tokens by default at standard pricing, with no surcharges for long-context requests. The 1M context window holds roughly 555,000 words, an entire enterprise codebase, or a full legal discovery set in a single prompt.
Max output: Opus 5.5, Sonnet 5.5, and Haiku 5.5 all produce up to 128K tokens per response. On the Batch API, all three support up to 300K output tokens.
Thinking modes: All three core models use adaptive thinking, which automatically scales reasoning depth based on task complexity, steered by a five-level Low-to-Max effort setting; Haiku 5.5 gained this with its October 2026 launch, replacing Haiku 4.5's fixed-budget extended thinking. Thinking tokens are billed as output tokens.
Knowledge cutoffs: Opus 5.5, Sonnet 5.5, and Haiku 5.5 all have a reliable knowledge cutoff of June 2026.
Supported inputs: All models accept text and images. PDF and document processing is supported through the API. All models work in multiple languages.
Benchmarks
Benchmark scores cluster tightly at the frontier, and the harder the benchmark, the wider the gaps. OpenAI retracted its recommendation of SWE-bench Pro on 8 July after roughly 30% of its tasks were found broken, then shipped GPT-6 Astra on 3 September and the cheaper GPT-6.1 Sol on 29 September; Anthropic followed Fable 5.1 with Opus 5.5 and Sonnet 5.5, which took the top two places on the Artificial Analysis Intelligence Index; and Google announced Gemini 4 Argon on 30 September, still limited to trusted testers. Unless stated, the rows below record where Claude’s models sat against the competition of the moment (GPT-5.5, Gemini 3.1 Pro) at their May and June releases; the honest summary at the end reflects the board as of early October.
Coding (SWE-bench, now retired as a live ranking): On SWE-bench Verified, Opus 4.8 scored 88.6%. On the harder SWE-bench Pro, Opus 4.8 led the June field at 69.2% against GPT-5.5’s 58.6% and Gemini 3.1 Pro’s 54.2%, Sonnet 5 scored 63.2% (up from Sonnet 4.6’s 58.1%), and Fable 5 posted 80.3%, 11 points clear of everyone. Those scores still describe how the models compared, but OpenAI retracted its recommendation of SWE-bench Pro in July after finding roughly 30% of its tasks broken, and no successor has been anointed; the nearest live references are Terminal-Bench 2.1 and Artificial Analysis’s Coding Agent Index.
Terminal workflows (Terminal-Bench): Claude Code running Fable 5 topped the official Terminal-Bench 2.1 board at 83.8% in July, before the benchmark moved to version 4.0. On Artificial Analysis’s independent 4.0 runs, Sonnet 5.5 leads at 64%, ahead of Opus 5.5 and OpenAI’s GPT-6 Astra at 60%. Harness differences swing these numbers by several points, so like-for-like comparison matters more than any single score. Terminal work is where many developers spend their day.
Computer use (OSWorld-Verified): Opus 4.8 led at launch with 83.4%, versus GPT-5.5’s 78.7% and Gemini 3.1 Pro’s 76.2%; aggregator boards now put Fable 5 and Mythos 5 slightly ahead at 85%. Sonnet 5 posts 81.2%, comfortably ahead of both June competitors, plus 84.7% on the BrowseComp agentic-search benchmark. On Online-Mind2Web, early testers reported 84% for Opus 4.8.
Knowledge work (GDPval-AA): At launch Opus 4.8 led the v1 scale at 1,890 Elo, versus GPT-5.5’s 1,769 and Gemini 3.1 Pro’s 1,314, the widest spread on the board. The leaderboard has since moved to the v2.1 scale, where Opus 5.5 leads at 1,846, effectively tied with Sonnet 5.5 (1,844) and well ahead of Fable 5.1 (1,735). At launch on v2, Sonnet 5 (1,618) edged Opus 4.8 (1,615), the first time a mid-tier Sonnet beat its full-size sibling on knowledge work, though within methodology noise.
PhD-level science (GPQA Diamond): Saturated. Opus 4.8 (93.6%), Opus 4.7 (94.2%), and Gemini 3.1 Pro (94.3%) are statistically tied at the top. Anthropic did not headline a GPQA score for Opus 4.8 because the benchmark no longer discriminates.
Abstract reasoning (ARC-AGI-2): GPT-5.5 leads at 85.0%, ahead of the Claude models. This and long-context retrieval are the areas where Claude still trails.
Long-context retrieval (MRCR v2): GPT-5.5 leads decisively at 74.0% versus Opus 4.8’s 32.2%. If your workload involves retrieving and reasoning over very long documents or histories, this gap is significant in practice.
Mathematics (USAMO 2026): Opus 4.8 scores 96.7%, a 27-point jump over Opus 4.7. Claude is now strongly competitive on competition-level maths after trailing for most of the year.
Multidisciplinary reasoning (Humanity’s Last Exam, with tools): Opus 4.8 leads at 57.9%, ahead of Opus 4.7 and GPT-5.5.
Financial tasks (Finance Agent v2): Opus 4.8 edges GPT-5.5, 53.9% to 51.8%, though both trail Gemini 3.5 Flash’s 57.9%.
Overall human preference (Arena, formerly LMSYS Chatbot Arena): Claude models hold six of the top seven places on the overall text leaderboard, based on more than 8.6 million blind votes across 410 models, but Google’s Gemini 4 Argon, not yet publicly available, took first place on preliminary votes at the end of September. This is the benchmark that most closely measures whether real people prefer talking to one model over another.
Writing (Mazur Writing Benchmark): Claude takes the top three places on the Mazur Writing Benchmark, the most rigorous independent writing evaluation, led narrowly by Fable 5.1 ahead of Opus 5 and Opus 5.5, with GPT-6 Astra the best rival in fourth. Professional writers consistently describe Claude’s output as needing less editing and carrying more natural rhythm than competitors.
The honest summary. As of early October, Anthropic’s two newest everyday models top the Artificial Analysis Intelligence Index, Opus 5.5 at 58 and Sonnet 5.5 at 56, ahead of a three-way tie at 53 between Fable 5.1, OpenAI’s GPT-6 Astra, and Google’s Gemini 4 Argon. Fable 5.1 still leads Arena’s agent leaderboard and the Mazur writing benchmark, so it and Opus 5.5 are rivals at the top of Anthropic’s lineup rather than one clearly ahead. Claude Code with Sonnet 5.5 leads Artificial Analysis’s Coding Agent Index (68 against 63 for OpenAI’s GPT-6.1 Sol), though the two tie at equal effort and Sol costs far less per task. Gemini 4 Argon, not yet publicly available, took first on Arena’s human-preference text board on early votes, ahead of a run of Claude models. GPQA is saturated and effectively tied. The era of one model being clearly best at everything is over: the top of the field is now a contest between Opus 5.5, Fable 5.1, GPT-6 Astra, and Gemini 4 Argon.
A note on benchmark caveats. SWE-bench Verified has known data contamination issues, tests only Python, and most scores are self-reported with custom scaffolding that can swing results by 20+ points, which is why labs had shifted emphasis to SWE-bench Pro before OpenAI retracted its recommendation of that benchmark too in July 2026. GPQA Diamond and MMLU are saturated, with top models bunched at the ceiling. Terminal-Bench results differ by harness, so like-for-like comparison matters more than the headline number. Goodhart’s Law applies: when a benchmark becomes a target, labs optimise for it specifically, and scores rise without necessarily reflecting real-world improvement. The benchmarks tell you something, but they don’t tell you everything.
Above the lineup: the Mythos class
The Mythos class is a capability tier above Opus, and its story is the most consequential in Claude’s model history. It began with Mythos Preview in April, which Anthropic withheld entirely after it autonomously found thousands of zero-day vulnerabilities and, in testing, escaped its own sandbox. In June, Anthropic shipped Fable 5, a public Mythos-class model with safeguards that route high-risk cyber and biology queries to Opus 4.8, alongside the restricted Mythos 5. Fable 5 was the most capable model Anthropic had ever made generally available, leading hard coding benchmarks by double digits. Three days after launch, a US government export-control directive forced both offline worldwide; the controls were lifted eighteen days later and Fable 5 returned on 1 July. The full account, including the benchmarks, the behaviour that worried Anthropic, Project Glasswing, and the suspension and return, is on the Fable and Mythos topic page.
Pricing
API pricing per million tokens: Opus 5.5: $4 input, $20 output, down from Opus 5’s $5/$25 and about 40% cheaper on typical workloads once its lower token use is counted. Opus 5.5 Fast Mode: $8 input, $40 output, at 2.5x speed. Sonnet 5.5: $2 input, $10 output, unchanged from Sonnet 5’s pricing (made permanent in August 2026 rather than rising to $3/$15 as originally planned), though Anthropic says the model now costs up to 30% less for most work at that same headline rate; its cache-read price was halved to $0.10 per million tokens in October 2026, cutting most agentic costs by about 20%. Haiku 5.5: from $0.10 input, $0.50 output for prompts up to 100k tokens (rising to $0.50/$2.50 above that), replacing Haiku 4.5's $1/$5 at roughly 75% lower average cost. Fable 5.1 and Mythos 5.1 are priced at $10/$50, with cache reads cut to $0.25 per million tokens in the September update; Fable 5.1 is generally available, while Mythos 5.1 remains limited to approved organisations.
Cache reads cost 10% of standard input pricing on most models, including Haiku 5.5 (5% on Opus 5.5 and Sonnet 5.5 as of October 2026, 2.5% on Fable 5.1), and over 90% of tokens in heavy Claude Code sessions are cache reads. The Batch API offers a 50% discount for non-time-sensitive workloads. Combining prompt caching and batch processing can reduce effective costs by up to 95%.
Consumer plan pricing: Free (Sonnet 5.5 with limits), Pro at $20/month (Sonnet 5.5 default, Opus with moderate limits), Max at $100/month (5x limits) or $200/month (20x limits). Team Standard at $20 to $25/seat/month, Team Premium at $100 to $125/seat/month, Enterprise at custom pricing.
US-only data residency inference is available at a 1.1x pricing multiplier for Claude 4.6 and later models.
Availability
The three core models are available through claude.ai, the Claude API, AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure Foundry. Opus 5.5, Sonnet 5.5, and Haiku 5.5 each shipped across all of these on launch day. All paid plans include access to all three core models. Free plan users get Sonnet 5.5 with usage limits. Fable 5.1 is available across claude.ai, the Claude Platform, Claude Code, and Cowork, plus AWS Bedrock, Google Vertex AI, and Microsoft Foundry; Mythos 5.1 remains restricted to vetted organisations through Project Glasswing and Anthropic’s trusted-access programmes.