Claude is Anthropic’s family of large language models, designed with a focus on safety, reliability, and extended context handling. For enterprise AI applications, Claude is one of the most widely deployed models via Amazon Bedrock, where it is available across multiple capability tiers - and, as of September 2026, also via Google Cloud, Microsoft Foundry, and Anthropic’s own Claude Platform on AWS.

Model Tiers

As of 8 October 2026, Anthropic’s current lineup spans five named models across four tiers (Haiku 5.5, Sonnet 5.5, Opus 5.5, and Fable 5.1 with its gated sibling Mythos 5.1), with Sonnet 5 and Haiku 4.5 still available as legacy models, all sharing the same API and the same general capability profile (tool use, long context, strong instruction following) at different price/performance points. Model IDs in the 5 generation are dateless pinned snapshots; Haiku 4.5’s is an alias that resolves to a dated snapshot, which matters if you pin versions in production.

Haiku 5.5 (claude-haiku-5-5) - Released on 7 October 2026, completing the 5.5 generation, and now the Haiku tier to start on. 1M-token context window, 128K maximum output (300K on the Message Batches API with the output-300k-2026-03-24 beta header), adaptive thinking with a default effort of medium, and a June 2026 reliable knowledge cutoff. It is the only current Claude model priced by prompt length: $0.10/$0.50 per MTok for prompts up to 100,000 tokens and $0.50/$2.50 above, with cache reads at $0.01 and $0.05. Anthropic says that makes it about 75% cheaper to run than Haiku 4.5 on average, and that its updated tokenizer uses slightly more tokens per task. It is the first Haiku model with the effort parameter. Retirement not sooner than 7 October 2027. See Claude Haiku 5.5 ships at $0.10/$0.50 .

Haiku 4.5 (claude-haiku-4-5, an alias for the pinned snapshot claude-haiku-4-5-20251001) - The fastest and most cost-efficient tier, with a 200K-token context window and a 64K maximum output. Best for high-volume, lower-complexity tasks: classification, short-form extraction, simple summarization, intent detection. At $1/$5 per million tokens against Sonnet 5’s $2/$10, Haiku costs exactly half Sonnet’s per-token price - a real saving, but a much narrower one than earlier Haiku generations offered against their contemporary Sonnet, so build the cost model on the current numbers rather than on the old order-of-magnitude intuition. Haiku 4.5 is also the odd one out in the lineup in three ways: released 15 October 2025, it is the only current model still on the previous tokenizer, the only one that does not support the effort parameter (it uses manual extended thinking with budget_tokens), and the one closest to end of life - Anthropic’s published floor is that it will not retire before 15 October 2026. No deprecation has been announced, and Anthropic commits to at least 60 days’ notice for publicly released models, so it cannot arrive unannounced - but Haiku 4.5 is the model in the current lineup to keep a migration plan for. Its knowledge cutoff (February 2025) is also by far the oldest on offer.

Sonnet (claude-sonnet-5) - The middle tier, balancing capability and cost, with a 1M-token context window and 128K maximum output. Anthropic released Claude Sonnet 5 on 30 June 2026 for agentic and automation workloads, at performance close to the prior-generation flagship, Opus 4.8. Its introductory pricing was scheduled to rise 50% on 1 September 2026; Anthropic cancelled that increase on 10 August 2026 and made the introductory rate the permanent standard price instead. A standing note to that effect now sits on Anthropic’s own pricing page, so this is settled rather than provisional. See Anthropic cancels the Sonnet 5 price rise . Sonnet 5 uses adaptive thinking steered by the effort parameter (default high); manual extended thinking and non-default temperature, top_p or top_k all return a 400. For most enterprise knowledge-worker tasks (document review, draft generation, data analysis, high-volume agentic automation), Sonnet 5 provides the best practical value. Claude Sonnet 5.5 shipped on 28 September 2026 as claude-sonnet-5-5, at the same $2/$10 per MTok as Sonnet 5, with cache reads at $0.20 and cache writes at $2.50. Anthropic says it runs more than 30% faster than Sonnet 5 and costs up to 30% less for most work, because it uses fewer tokens per task as well as costing less per token. It ships safety classifiers aimed at distillation attacks, cybersecurity safeguards comparable to Opus 5.5, and the same biology safeguards as Sonnet 5, and it is available on the Claude Platform, AWS, Google Cloud and Microsoft Azure with a zero data retention option. Anthropic did not publish its context window or maximum output in the release post. Claude Haiku 5.5, announced in the same 22 September statement, has still not shipped. See Claude Sonnet 5.5 lands at $2/$10, faster and cheaper than Sonnet 5 . For new Sonnet-tier work, start on Sonnet 5.5.

Opus (claude-opus-5-5) - The general-purpose flagship for long-running agentic coding and knowledge work, at 1M context and 128K maximum output (300K on the Message Batches API behind the output-300k-2026-03-24 beta header). Claude Opus 5.5 reached general availability on 22 September 2026 (see Claude Opus 5.5 launches at $4/$20, with thinking always on ), the first model in Anthropic’s new Claude 5.5 family, and it is where Anthropic’s own documentation now tells you to begin: “If you’re unsure which model to use, start with Claude Opus 5.5 for most workloads.” Anthropic says it “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5” - the 40% figure combines a 20% per-token price cut ($4/$20 per MTok, against Opus 5’s $5/$25) with fewer tokens used per task on Anthropic’s own tests, so treat it as a vendor estimate to re-measure on your workload. Cache reads cost $0.20/MTok (5% of input, a multiplier specific to Opus 5.5). Its reliable knowledge cutoff is June 2026, the default effort is medium (not high as on Fable 5.1 and Sonnet 5), and adaptive thinking is always on: thinking: {"type": "disabled"} and manual "enabled" both return a 400. Forced tool_choice ("any" or a named "tool") also returns a 400, as on Fable 5.1; so do non-default temperature, top_p or top_k and a prefilled final assistant turn (both already rejected on recent Opus models); and on the Claude API and Google Cloud computer use requires computer_toolset_20260801 (Bedrock still accepts computer_20251124). Anthropic’s migration guide lists these changes for teams moving off Opus 5. Opus 5.5 is also the first Opus model to launch with Fable 5.1-class safeguards: a biology classifier alongside the cybersecurity one, and most cybersecurity tasks are re-routed to Opus 4.8 (Anthropic says routine find-and-fix of bugs in your own code is still handled). Its retirement floor is 22 September 2027. Anthropic’s Opus 5.5 model pages do not list the 30-day-retention requirement that applies to Fable 5.1 and Mythos 5.1; AWS states that Opus 5.5 on Amazon Bedrock has zero data retention support by default, but Anthropic has not published an equivalent statement for its first-party API, so confirm ZDR eligibility with your account team before moving a ZDR workload there.

Claude Opus 5 (claude-opus-5, GA 24 July 2026) is now previous generation, superseded by Opus 5.5: Anthropic’s models overview has moved it to “Legacy models (still available)”, and the deprecations table still lists it Active with a retirement floor of 24 July 2027. Anthropic documents it as available under zero-data-retention arrangements. Adaptive thinking on Opus 5 can still be disabled at effort high or below, which is the practical reason some integrations may stay on it while they migrate.

Fable (claude-fable-5-1) - Anthropic’s most capable widely released model, positioned above Opus for the most demanding programming and long-horizon knowledge work, also at 1M context and 128K maximum output. Claude Fable 5.1 reached general availability on 1 September 2026, at the same headline per-token price as its predecessor, Fable 5, but with prompt-cache-read pricing cut 75% (from $1 to $0.25 per million tokens) - a change that disproportionately benefits agentic, tool-heavy workloads that repeatedly re-read cached context. Being the most capable model does not make it the default: Anthropic’s guidance is to start on Opus 5.5 and reach for Fable 5.1 “for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short.” With Opus 5.5 priced at 40% of Fable 5.1’s per-token rate and claimed to match it on most work, the case for Fable 5.1 now has to be made on evals rather than assumed. See Claude Fable 5.1 reaches general availability, Mythos 5.1 stays gated .

Mythos (claude-mythos-5-1) - The same capabilities and specifications as Fable 5.1, offered to cybersecurity defenders and life-sciences researchers who need capabilities the general-availability safeguards would otherwise block. Unlike every other tier above, Mythos 5.1 is not self-serve: access runs through Anthropic’s vetted-access programs and, at launch, is limited to a set of US organizations, with expansion planned through government coordination with domestic and international partners. Which program actually reaches Mythos 5.1 today is worth reading carefully. The Life Sciences Verification Program (LSVP) - developed in partnership with the US government - is the one that grants Mythos 5.1 access now. The Cyber Verification Program (CVP), as of the 1 September announcement, still provides access only to certain Opus- and Sonnet-class models with reduced cyber safeguards for defensive security work; Anthropic says the program “will also include access to Claude Mythos-class models” in the near future, but that is a stated plan, not a shipped capability. A defensive-security team should not assume CVP enrolment currently yields a Mythos 5.1 key. Anthropic’s docs direct prospective users to their Anthropic, AWS, or Google Cloud account team. Pricing is identical to Fable 5.1 ($10/$50 per MTok, $0.25/MTok cache reads) for the organizations that qualify. One platform difference is easy to miss: Mythos 5.1 is listed on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, but not on Claude Platform on AWS, which Fable 5.1 does support.

How much lighter Mythos’s safeguards are is less clearly documented than the access program. For the previous generation Anthropic said it plainly: Fable 5 “includes safety classifiers that can decline requests, and Claude Mythos 5 does not.” For 5.1 the migration guide files safety classifiers under where the two models diverge and describes only Fable 5.1’s behavior, while still telling Mythos 5.1 integrators to handle stop_reason: "refusal" and read stop_details.category. The lighter-safeguard framing is directionally right, but Anthropic has not restated it for this generation, so do not design around an assumption that Mythos 5.1 never refuses.

Deployment constraints specific to Fable 5.1 and Mythos 5.1. Both models require 30-day data retention and are not available under zero-data-retention arrangements unless Anthropic expressly authorizes it; a request from an organization or workspace without 30-day retention returns a 400 invalid_request_error. Both are designated Covered Models, neither is supported on Priority Tier, text they generate carries an Anthropic watermark, and media produced by the code execution tool carries C2PA Content Credentials. For a regulated buyer on a ZDR contract, that is the binding constraint on the entire top tier - Opus 5, by contrast, is available under ZDR. Three API behaviors also changed at 5.1 relative to Fable 5: forced tool_choice ("any" or "tool") now returns a 400, thinking blocks are readable only by the model that produced them or a newer one, and on Fable 5.1 editing an earlier turn invalidates thinking blocks (Mythos 5.1 does not run that check).

Anthropic also continues to serve the previous generation for existing integrations: Opus 5 (24 July 2026), Fable 5 and Mythos 5 (both released 9 June 2026), Opus 4.8, 4.7, 4.6 and 4.5, and Sonnet 4.6 and 4.5. “Legacy” is a defined lifecycle term - no further updates, possible deprecation later - but it is not deprecation: the models overview files Fable 5, Opus 5, the four Opus 4.x releases and the two Sonnet 4.x releases under “Legacy models (still available)”, their model pages read Active (legacy), and the model-deprecations table that actually governs support commitments lists every one of them as Active with no deprecation announced. Mythos 5 appears in neither list, because invite-only models are omitted from both; its own model page reads Active (invite only), with the same 9 June 2027 retirement floor as Fable 5. Most of those have retirement floors in 2027, but two are close: Sonnet 4.5 may retire no sooner than 29 September 2026 - four days from this update - and Opus 4.5 no sooner than 24 November 2026. Neither has a formal deprecation notice yet, and Anthropic commits to at least 60 days’ notice, so neither can actually retire on its floor date; but they are the first two to plan off. Two differences are worth knowing before leaving a Fable 5 integration in place - its cache reads cost $1.00/MTok rather than $0.25, and it is supported on Priority Tier, which Fable 5.1 is not. Further back, Opus 4.1 retired on 5 August 2026 and Opus 4 and Sonnet 4 on 15 June 2026; those are now reachable only through the partner platforms that still list them. Anthropic’s published replacement for all three is the prior generation, not the current flagship - Opus 4.1 and Opus 4 point at claude-opus-4-8 and Sonnet 4 at claude-sonnet-4-6, because each deprecation was announced before Opus 5 shipped. Treat those notices as the minimum safe landing spot, not the destination: for new work, Sonnet 5, Opus 5.5, and Fable 5.1/Mythos 5.1 are the recommended defaults at their respective tiers.

Key Strengths

Long context window - Sonnet 5, Opus 5.5, Opus 5, Fable 5.1, and Mythos 5.1 all support context windows up to 1 million tokens - roughly 555,000 words on the current tokenizer (introduced with Opus 4.7, which Anthropic says produces about 30% more tokens per unit of text than the previous one; its own word-count figures, 750,000 words per million tokens before and 555,000 after, imply closer to 35%) - billed at standard per-token rates across the full window, with a 128K maximum output. Opus 5.5, Opus 5 and Sonnet 5 can go to 300K output on the Message Batches API behind the output-300k-2026-03-24 beta header; Fable 5.1 and Mythos 5.1 cannot. Haiku 4.5, which still uses the previous tokenizer, supports 200K tokens (roughly 150,000 words) and 64K of output. For document generation and long-form synthesis the output ceiling binds long before the context window does, so size both. This means entire books, large codebases, or extensive document sets can be processed in a single call without chunking. For RAG pipelines where traditional chunking limits retrieval quality, Claude’s long context can process entire source documents directly.

Instruction following - Claude reliably follows complex, structured instructions, including multi-part JSON output schemas, tone and style constraints, and conditional logic instructions. This makes it well-suited for applications with strict output format requirements.

Reduced hallucination on constrained tasks - When instructed to answer only from provided context and to acknowledge when information is not present, Claude follows these constraints reliably. This is critical for enterprise knowledge base applications where accurate sourcing matters.

Code tasks - Claude performs well on code generation, debugging, and documentation tasks, including less common languages and frameworks. Fable 5.1 in particular is positioned by Anthropic specifically around programming and long-running agentic problem-solving, with reported benchmark gains over Fable 5 on coding- and tool-use-heavy evaluations.

Access Options

Claude is available through:

  • Anthropic API - Direct API access with usage-based pricing
  • Amazon Bedrock - Access via Bedrock API with AWS IAM, VPC, and compliance controls; data stays within your AWS account; pricing is partner-operated and set independently by AWS
  • Claude Platform on AWS - Anthropic’s own AWS Marketplace listing, billed in Claude Consumption Units (CCU) at Anthropic’s standard per-model rates, so usage shows as a single line item on your AWS bill
  • Google Cloud - via the Gemini Enterprise Agent Platform (formerly Vertex AI); see Google Vertex AI for what changed in that rebrand
  • Microsoft Foundry - billed through the Azure Marketplace, also using Claude Consumption Units at standard first-party rates
  • Claude.ai - Web and API interface for direct use and team collaboration features

Coverage is not uniform across the lineup, so check it per model rather than per vendor. Opus 5.5, Sonnet 5, Haiku 4.5 and Fable 5.1 are listed on all of these paths - for Opus 5.5 the model page, the models comparison table and the 22 September API release note agree, including Claude Platform on AWS. AWS’s launch notes add that Opus 5.5 is also available in AWS GovCloud (US) and that on Bedrock it comes with zero data retention (ZDR) support by default. Mythos 5.1’s availability row omits Claude Platform on AWS. The previous-generation Opus 5 remains the awkward case, because Anthropic’s own pages disagree: its model page lists only the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, but the API release note dated 24 July 2026 says Opus 5 is available on Claude Platform on AWS. For new work on that path, Opus 5.5 removes the question.

For enterprise deployments requiring data residency controls, compliance certifications (SOC 2, HIPAA BAA), and IAM integration, Bedrock or Claude Platform on AWS is typically the preferred access path - with one caveat worth pricing in: that first-party data residency is not free: setting inference_geo: "us" applies a 1.1x multiplier to every token category (input, output, cache writes and cache reads) on Claude 4.6 and later models, on the Claude API and Claude Platform on AWS, and on Foundry’s US Data Zone Standard deployments. An organization whose cloud commitment is specifically with Google Cloud is, as of this writing, the one procurement path that reaches Claude but not OpenAI’s frontier models, since OpenAI does not offer GPT-5.6 or newer through Vertex AI’s Model Garden.

Common Enterprise Use Cases

  • Document processing - Contract review, policy analysis, report summarization
  • Customer service - Intelligent response drafting, inquiry triage, escalation detection
  • Code assistance - Review, generation, refactoring, documentation
  • Knowledge management - Q&A over internal documentation, knowledge base search and synthesis
  • Data analysis - Interpreting structured data outputs, generating narrative from metrics, analyzing reports

Pricing

Current Claude API pricing, per million tokens (MTok), fetched directly from Anthropic’s pricing documentation:

ModelInputCache write (5m / 1h)Cache readOutputContext / max output
Claude Haiku 5.5 (prompts up to 100K tokens)$0.10$0.125 / $0.20$0.01$0.501M / 128K
Claude Haiku 5.5 (prompts over 100K tokens)$0.50$0.625 / $1.00$0.05$2.501M / 128K
Claude Haiku 4.5 (previous generation)$1.00$1.25 / $2.00$0.10$5.00200K / 64K
Claude Sonnet 5.5$2.00$2.50 / $4.00$0.10$10.001M / 128K
Claude Sonnet 5$2.00$2.50 / $4.00$0.20$10.001M / 128K
Claude Opus 5.5$4.00$5.00 / $8.00$0.20$20.001M / 128K
Claude Opus 5 (previous generation, superseded by Opus 5.5)$5.00$6.25 / $10.00$0.50$25.001M / 128K
Claude Fable 5.1$10.00$12.50 / $20.00$0.25$50.001M / 128K
Claude Mythos 5.1 (verified orgs only)$10.00$12.50 / $20.00$0.25$50.001M / 128K

Notes:

  • Sonnet 5.5’s cache reads were $0.20 at launch and were halved to $0.10 on 7 October 2026, alongside the Haiku 5.5 release. Its 1-hour cache write ($4.00) and its 1M / 128K limits, which the release post did not state, are now published on Anthropic’s pricing page and models overview.
  • Haiku 5.5 is the only current model whose price depends on prompt length: a request over 100,000 prompt tokens is billed at the higher row for the whole request. Every other 4.6-and-later model bills the full 1M window at its standard rate.
  • Sonnet 5’s $2/$10 rate was introductory pricing through 31 August 2026, scheduled to rise 50% to $3/$15; Anthropic cancelled that increase on 10 August 2026 and made $2/$10 the permanent standard rate.
  • Cache writes cost 1.25x the base input rate for the 5-minute TTL and 2x for the 1-hour TTL, on every model. Cache reads cost 10% of the base input rate on every model except Fable 5.1 and Mythos 5.1, which use a sharper 2.5% multiplier, and Opus 5.5 and Sonnet 5.5, which use 5% ($0.20/MTok and $0.10/MTok). Anthropic notes cache reads make up the majority of agentic and coding costs, so on cache-heavy workloads Opus 5.5’s saving over Opus 5 ($0.20 vs $0.50) is larger than the 20% headline per-token cut. Caching pays back only when the write premium is amortized over enough reads, so compute it from both columns rather than the read column alone.
  • The Batch API discounts input and output tokens 50% on every model above (asynchronous, non-latency-sensitive workloads only).
  • Modifiers that do not appear in the table but do appear on the bill: inference_geo: "us" adds a 1.1x multiplier to every token category on Claude 4.6 and later (Claude API, Claude Platform on AWS, and Foundry US Data Zone Standard deployments); fast mode, a research preview available on Opus 5.5, Opus 5 and Opus 4.8 through the Claude API only, bills at $8/$40 per MTok on Opus 5.5 (Anthropic claims up to 2.5x faster output) and $10/$50 on Opus 5 and Opus 4.8 - double each model’s standard rate; and Claude Managed Agents adds a session-runtime charge of $0.08 per session-hour on top of tokens.
  • Tool pricing is separate from tokens: web search is $10 per 1,000 searches, web fetch is free, and code execution is free when paired with web search or web fetch, otherwise $0.05 per container-hour beyond 1,550 free hours per organization per month.
  • Amazon Bedrock and Google Cloud are partner-operated: pricing is set independently by AWS and Google and can differ from the first-party rates above - check Bedrock pricing or Google Cloud pricing directly.
  • Claude Platform on AWS and Claude in Microsoft Foundry are Anthropic-operated and bill at the same rates as the table above, converted to Claude Consumption Units ($0.01/CCU) for marketplace invoicing.
  • Pricing changes over time - always check the current Anthropic pricing page before building cost models for production workloads.

Recent Developments

  • 28 May 2026 - Anthropic raised a $65 billion Series H at a $965 billion post-money valuation, alongside the Opus 4.8 launch. See Anthropic raises $65 billion at a $965 billion valuation .

  • 9 June 2026 - Claude Fable 5 and Claude Mythos 5 released, opening a new top tier above Opus. Both are still served today; Anthropic’s docs file Fable 5 under legacy models and give Mythos 5 an invite-only page of its own, but the lifecycle table still marks Fable 5 Active, and both carry a retirement floor of 9 June 2027.

  • 12 June - 1 July 2026 - A US national-security order issued on 12 June required Anthropic to disable Fable 5 and Mythos 5 for foreign nationals, and Anthropic disabled both worldwide. The Commerce Department lifted the order on 30 June; Mythos 5 had already returned on 26 June for a vetted set of US organizations that operate and defend critical infrastructure - Anthropic published no count, though press reports put it above 100 - and Fable 5 returned worldwide on 1 July after Anthropic shipped a classifier blocking the reported jailbreak. The staggered restoration is the origin of the vetted-access pattern Mythos still runs on. See Why the US restricted Anthropic’s Fable 5 and Mythos 5 and US lifts export controls on Fable 5 and Mythos 5 .

  • 30 June 2026 - Claude Sonnet 5 released for agentic and automation workloads. See Anthropic releases Claude Sonnet 5 .

  • 24 July 2026 - Claude Opus 5 reached general availability, succeeding Opus 4.8 as the flagship and Anthropic’s stated starting point for most workloads. Fast mode was removed from Opus 4.7 the same day (speed: "fast" now errors there).

  • 5 August 2026 - Claude Opus 4.1 retired on Anthropic-operated platforms, with Anthropic’s deprecation notice directing users to Opus 4.8 (the notice was published 5 June 2026, before Opus 5 shipped); Opus 4 and Sonnet 4 had already retired on 15 June 2026, pointed at Opus 4.8 and Sonnet 4.6 respectively.

  • 10 August 2026 - Sonnet 5’s introductory $2/$10 pricing was made permanent, cancelling a scheduled 50% increase. See Anthropic cancels the Sonnet 5 price rise .

  • 19 August 2026 - The Files API and Agent Skills left beta on the Claude API (the files-api-2025-04-14 and skills-2025-10-02 headers are no longer required), computer use went GA as computer_toolset_20260801, and a browser_toolset_20260801 shipped alongside it.

  • 1 September 2026 - Claude Fable 5.1 and Claude Mythos 5.1 reached general availability, with Mythos 5.1 restricted to verified organizations. Anthropic reports roughly 25% lower cost on typical workloads and up to 45% on complex agentic tasks, driven mainly by the cache-read cut, with benchmark gains over Fable 5 on Terminal-Bench-Science 0.1 (52.6% vs 24.7%), Terminal-Bench 4.0 (55.8% vs 42.0%) and OSWorld 2.0 strict (41.7% vs 36.1%). Text watermarking, C2PA content credentials and the 30-day retention requirement shipped the same day. See Claude Fable 5.1 reaches general availability, Mythos 5.1 stays gated .

  • 17 September 2026 - Claude Code relaunched Projects as multi-agent cloud threads: a coordinator splits a goal into parallel Claude Code cloud sessions, each on its own git branch and copy of the repo, in beta for select Pro and Max subscribers. See Claude Code relaunches Projects as coordinated parallel cloud threads and Claude Code .

  • 22 September 2026 - Claude Opus 5.5 reached general availability on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry, at $4/$20 per MTok, superseding Opus 5 as Anthropic’s recommended starting model. Anthropic said Claude Sonnet 5.5 and Haiku 5.5 would follow “in the coming weeks” (Sonnet 5.5 arrived six days later, Haiku 5.5 on 7 October) and that its Cyber Verification Program will be expanded to Opus 5.5 with three tiers of increasingly permissive access, including access to Mythos models. The same day, fast mode came to Opus 5.5 and tools became definable inside mid-conversation system messages (beta, inline-tools-2026-09-15).

  • 28 September 2026 - Claude Sonnet 5.5 released as claude-sonnet-5-5 at $2/$10 per MTok, unchanged from Sonnet 5, on the Claude Platform, AWS, Google Cloud and Microsoft Azure. Anthropic reports it runs more than 30% faster than Sonnet 5 and costs up to 30% less for most work, and reports Terminal-Bench 4.0 at 70.6% against Sonnet 5’s 10.3%, CursorBench 4.0 at 55.5%, FrontierCode 1.1 at 46.2% on Max effort, GDPval-AA v2.1 at 1844, OSWorld 2.1 partial at 80.1% and Chartography without tools at 61.6%. All of those are vendor-reported and none have been reproduced independently; the Terminal-Bench gap in particular is large enough that the harness or scoring most likely changed alongside the model. See Claude Sonnet 5.5 lands at $2/$10, faster and cheaper than Sonnet 5 .

  • 23-24 September 2026 - Cache diagnostics left beta on the Claude API (23 September). From 24 September, refusals that arrive before any output are billed again when stop_details.category is "bio", "frontier_llm" or "reasoning_extraction"; other pre-output refusals remain unbilled.

  • 7 October 2026 - Claude Haiku 5.5 released as claude-haiku-5-5 at $0.10/$0.50 per MTok up to 100K prompt tokens and $0.50/$2.50 above, on the Claude Platform, AWS, Google Cloud and Microsoft Azure. The same day Sonnet 5.5 cache reads were halved to $0.10 per MTok, Max 5x and Max 20x plans gained $100 and $200 a month in API credits (Team up to $500 pooled), and the Python and TypeScript SDKs added beta computer-use and browser-use support. Anthropic’s benchmark figures (OSWorld 2.1 offline subset 72.4%, Terminal-Bench 4.0 39.2%) are vendor-reported. See Claude Haiku 5.5 ships at $0.10/$0.50 .

As of 8 October 2026, the 5.5 generation is complete: Haiku 5.5, Sonnet 5.5 and Opus 5.5, with Fable 5.1 above them. Anthropic’s models overview now lists Haiku 4.5 and Sonnet 5 as legacy models that are still available. One retirement floor falls within the next week: Haiku 4.5 (15 October 2026), not yet formally deprecated, so plan the move to Haiku 5.5 now. Sonnet 4.5 passed its floor on 29 September 2026 without a deprecation notice, so it remains active.

For how this lineup compares to OpenAI’s models on cost, capability, and access constraints, see Claude vs ChatGPT and OpenAI vs Anthropic .

The Claude Product Family

Beyond the models and the chat apps, Anthropic ships several products built on the same Claude models. They differ mainly in where they run and what they produce.

Where it livesWhat it is forPlan needed
Claude CodeTerminal and IDEsEditing, running, and shipping codePro, Max, Team Premium, or API
Claude Designclaude.ai (Anthropic Labs)Designing UI and documents as HTMLPro, Max, Team, Enterprise
Claude CoworkClaude app (desktop, and since September 2026 web and mobile)Autonomous multi-step knowledge workPaid plans (Pro and Max first for the merged app)
Claude apps and APIWeb, mobile, desktop, APIChat, analysis, building on the modelFree and up

For teams, Anthropic also offers Claude Tag , a shared agent that lives inside Slack channels.

Sources

  1. Anthropic, “Pricing” (fetched 25 September 2026 - per-model input, cache-write, cache-read and output rates, the permanent Sonnet 5 pricing note, the 0.05x Opus 5.5 cache-read multiplier, tokenizer note, batch rates, fast-mode pricing including Opus 5.5 at $8/$40, the inference_geo multiplier, Managed Agents session runtime, and tool pricing): https://platform.claude.com/docs/en/about-claude/pricing
  2. Anthropic, “Models overview” (fetched 25 September 2026 - comparison table with Opus 5.5 replacing Opus 5, the legacy-models list, per-platform model IDs, retirement floors, knowledge cutoffs, word-count conversions): https://platform.claude.com/docs/en/about-claude/models/overview
  3. Anthropic, “Model deprecations” (fetched 25 September 2026 - lifecycle table including the Opus 5.5 and Opus 5 retirement floors, retirement dates for Opus 4.1, Opus 4 and Sonnet 4, parameter deprecations): https://platform.claude.com/docs/en/about-claude/model-deprecations
  4. Anthropic, “API release notes” (fetched 25 September 2026 - dated entries through 24 September 2026, including the 22 September Opus 5.5 launch and breaking changes, the 24 July Opus 5 entry and the 10 August Sonnet 5 pricing change): https://platform.claude.com/docs/en/release-notes/api
  5. Anthropic, model overview pages for Fable 5.1 , Mythos 5.1 , Opus 5 , Sonnet 5 , Haiku 4.5 and Fable 5 , all fetched 9 September 2026 - release dates, context and output limits, per-platform availability, and the Fable 5 statement that Mythos 5 runs no safety classifiers.
  6. Anthropic, “Claude Fable 5.1 migration guide” (fetched 9 September 2026 - shared baseline settings, the 30-day retention and Covered Model requirements, Priority Tier exclusion, and the safety-classifier divergence): https://platform.claude.com/docs/en/models/fable-5-1/migration-guide
  7. Anthropic, “Introducing Claude Fable 5.1 and Claude Mythos 5.1” (1 September 2026 - CVP and LSVP program names, the US-organizations-only scope, the cache-read cut, and the benchmark figures): https://www.anthropic.com/claude-fable-and-mythos-5-1
  8. Anthropic, “Claude Sonnet 5” (30 June 2026): https://www.anthropic.com/news/claude-sonnet-5
  9. Anthropic, “Anthropic raises Series H at a $965B post-money valuation” (28 May 2026): https://www.anthropic.com/news/series-h
  10. Anthropic, “Project Glasswing”: https://anthropic.com/glasswing . Note that as of 9 September 2026 this page still describes the April 2026 Mythos Preview launch and a $25/$125 per-MTok post-preview price that the pricing page and the Mythos 5.1 model page both contradict; the CVP/LSVP names and the US-only scope come from source 7 instead.
  11. Anthropic, “Introducing Claude Sonnet 5.5” (28 September 2026, fetched 3 October 2026 - $2/$10 pricing, $0.20 cache reads, $2.50 cache writes, the 30% faster and up to 30% cheaper claims, the benchmark table, the anti-distillation classifiers and the platform list): https://www.anthropic.com/claude-sonnet-5-5
  12. Anthropic, “Introducing Claude Opus 5.5” (22 September 2026 - “Fable 5.1 level” and 40%-lower-cost claims, the Sonnet 5.5 and Haiku 5.5 “coming weeks” statement, cybersecurity re-routing to Opus 4.8, and the Cyber Verification Program expansion): https://www.anthropic.com/claude-opus-5-5
  13. Anthropic, “Claude Opus 5.5” model overview and “What’s new in Claude Opus 5.5” (fetched 25 September 2026 - platforms, cache pricing, default effort, breaking changes, safeguard categories): https://platform.claude.com/docs/en/models/opus-5-5/overview and https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5
  14. AWS, “Claude Opus 5.5 is now available on AWS” (22 September 2026 - Bedrock and Claude Platform on AWS availability, ZDR support by default on Bedrock): https://aws.amazon.com/about-aws/whats-new/2026/09/claude-opus-5-5-aws/ and “Claude Opus 5.5 is now available on AWS GovCloud (US)” (22 September 2026): https://aws.amazon.com/about-aws/whats-new/2026/09/claude-opus-5-5-aws-govcloud/
  15. The Verge, “Claude Code relaunches Projects to manage multiple AI agents in the cloud” (17 September 2026): https://www.theverge.com/ai-artificial-intelligence/997134/anthropic-claude-code-projects
  16. This wiki, Claude vs ChatGPT: Constraints Before Features and OpenAI vs Anthropic , cross-verified against Anthropic’s pricing and model pages on 9 September 2026.
  17. Anthropic, “Projects redesigned: from folder to conversation” (17 September 2026): https://claude.com/blog/projects-redesigned
  18. Anthropic, “Migrating to Claude Opus 5.5” (fetched 25 September 2026 - sampling-parameter, prefill and forced-tool-choice 400s): https://platform.claude.com/docs/en/models/opus-5-5/migration-guide
  19. Anthropic, “Introducing Claude Haiku 5.5” (7 October 2026, fetched 8 October 2026 - two-tier pricing, the 75% cheaper claim, effort levels, the Sonnet 5.5 cache-read cut, plan API credits and the benchmark table): https://www.anthropic.com/claude-haiku-5-5 . Anthropic’s pricing page and models overview , both fetched 8 October 2026, confirm the Haiku 5.5 and Sonnet 5.5 rows, the 1M / 128K limits, the June 2026 cutoff and the legacy status of Haiku 4.5 and Sonnet 5.

Three things on this page are deliberately hedged because Anthropic’s own sources do not settle them: whether the previous-generation Opus 5 is orderable on Claude Platform on AWS (its model page says no, source 4 says yes), whether Mythos 5.1 runs safety classifiers (stated plainly for Mythos 5, not restated for 5.1), and whether Opus 5.5 is available under zero-data-retention arrangements on Anthropic-operated platforms (AWS says it is on Bedrock; Anthropic’s model pages neither require 30-day retention nor state ZDR eligibility). The “40% cheaper than Opus 5” and “Fable 5.1-level” claims for Opus 5.5 are Anthropic’s own measurements.

Further reading