Claude Opus: the full family history

Claude Opus is Anthropic's highest-capability tier. Every release from Opus 3 to Opus 5, what changed, which versions still work, and which have been retired.

Opus is Anthropic's high-capability Claude tier, and the current member is Claude Opus 5, released 24 July 2026. Five Opus versions are still callable in August 2026 - Opus 4.5, 4.6, 4.7, 4.8 and Opus 5 - while Opus 4 and Opus 4.1 have been retired and now return an API error.

This page is the archive and the map for the whole tier: every release in order, what changed between them, which ones you can still call, and where Opus fits now that Claude Fable 5 sits above it in the lineup.

#Which Claude Opus model should you use in August 2026?

Opus 5, unless you have a specific reason not to. Anthropic's own guidance in the model docs is to "start with Claude Opus 5 for complex agentic coding and enterprise work", and every earlier Opus version is priced identically at $5/$25 per million tokens from Opus 4.5 onward - so there is no cost saving in staying behind. The reasons to pin an older ID are behavioural: a prompt suite tuned against Opus 4.7 or Opus 4.6 that you have not yet re-evaluated, or a dependency on a parameter that later versions reject.

The one genuine trade-off is upward, not backward. Fable 5 is positioned above Opus 5 for long-horizon agent work and costs $10/$50 - double. Opus 5 also has substantially more generous rate limits: at the Start tier, 2,000,000 input tokens per minute against Fable 5's 500,000. If you are choosing across tiers rather than within one, the Haiku vs Sonnet vs Opus comparison works it through on cost per task.

#Every Claude Opus version, in order

Release dates, specifications and status for the whole tier. Prices are the per-million-token API rates listed by Anthropic; where a figure is not present in Anthropic's current documentation it is marked as such rather than reconstructed.

^* Figures marked with an asterisk are the model's specifications at launch, recorded here as history: Opus 4 and Opus 4.1 are retired, and Anthropic's current model pages no longer publish specifications for them. The fuller record is on the Claude Opus 4 page.

#What actually changed between Opus generations

Most of the visible movement in this tier since late 2025 has been in price, context and the thinking controls, rather than in headline capability claims. Anthropic publishes results for the newest models as chart images and relative statements rather than numeric tables, so the honest version of this history is the one built from specifications and documented behaviour.

#Opus 4 to Opus 4.1 (2025)

An incremental revision at the same $15/$75 price. Anthropic's announcement gave Opus 4.1 a SWE-bench Verified score of 74.5%, against 72.5% for Opus 4 at launch. Both are now historical: SWE-bench Verified is no longer reported by frontier vendors, and both IDs are retired.

#Opus 4.1 to Opus 4.5 - the price cut

The single largest change in the tier's history is commercial, not technical. Opus went from $15/$75 to $5/$25 per million tokens - a two-thirds reduction that put Opus within reach of workloads that had previously been forced onto Sonnet. Opus 4.5 also introduced the effort parameter in public beta. Its model ID carries the snapshot date 20251101 while the launch note is dated 24 November 2025 - a trap for anyone reconstructing the timeline from IDs alone.

#Opus 4.5 to Opus 4.6 - the million-token window

Opus 4.6 quadrupled context from 200k to 1M tokens and doubled maximum output to 128k. At launch in February 2026 the 1M window was a beta with long-context pricing above 200k input tokens; that premium has since gone, and on Claude 4.6 and later the full window bills at standard rates with no beta header. Compaction - server-side context summarisation - also arrived in beta here.

#Opus 4.6 to Opus 4.7 - a new tokenizer and stricter parameters

Opus 4.7 shipped a new tokenizer that produces roughly 30% more tokens for the same text, which means a like-for-like migration raises your bill even though the sticker price did not move. It also began returning a 400 error for non-default temperature, top_p and top_k values, and added task budgets in beta.

#Opus 4.7 to Opus 4.8 to Opus 5

Opus 4.8 added mid-conversation system messages and, per Anthropic, is "approximately four times less likely than Opus 4.7 to overlook flaws in code" - a relative claim with no published figure behind it. Opus 5 turned thinking on by default with effort defaulting to high, cut the minimum cacheable prompt to 512 tokens (Opus 4.5 and 4.6 require 4,096), and moved to a May 2026 knowledge cutoff. The mechanics of that thinking ladder are covered on extended and adaptive thinking.

#Where Opus sits in the lineup now

Opus was the top of the range from March 2024 until June 2026. It is not any more. Fable 5 took the capability crown, and Opus 5 was repositioned as the general default for complex agentic coding and enterprise work - the model most organisations should reach for first, with Fable reserved for jobs where a failure costs more than the tokens.

Two facts make that positioning less obvious than it looks. Opus 5's knowledge cutoff of May 2026 is newer than Fable 5's January 2026, so the cheaper model is the better-informed one about recent events. And on the independent Artificial Analysis Intelligence Index, Opus 5 holds the top spot at 63, while arena.ai's Elo leaderboard puts Fable 5 first. Which ranking you weight is itself a judgement call.

#Who the Opus tier is for

Use Opus when the task is long, structurally complex, or expensive to get wrong: multihour autonomous coding agents, large-scale refactoring, systems engineering, vision-heavy workflows and computer use. Use Sonnet 5 instead for the large middle of everyday work - it is a fifth of the output price and, on Anthropic's own recommendation, the sensible default. The family history for that tier is on the Sonnet page, and the small, fast tier is covered on the Haiku page.

Subscribers should know that Opus is metered separately. On Pro, Max, Team and Enterprise plans, Opus models are tracked on their own weekly counter alongside the shared allocation that chat, Claude Code, Cowork, Design and Claude for Excel all draw from. That is why heavy Opus users hit a wall while their general usage still has headroom - the mechanics are set out on the pricing page, and the whole current range is indexed on the Claude models page.

#Frequently asked questions