Claude Opus 5: Near-Fable Intelligence at Half the Cost

Claude Opus 5 went generally available on July 24, 2026. Per Anthropic's launch announcement, it delivered greatly improved performance for the same cost as its predecessor, Opus 4.8, and got close to the intelligence of the frontier Claude Fable 5 at substantially lower cost per task. The API identifier is claude-opus-5.

The headline is the price. Opus 5 launched at the identical $5 in / $25 out per MTok that Opus 4.8 charged, so the entire performance gain arrives at zero extra cost. It also ships with adjustable effort: you can push it toward maximum intelligence on a hard feature, or dial it down to conserve tokens for faster, cheaper runs on routine work. That flexibility made it a genuine default rather than a top-shelf model you ration, until Opus 5.5 took the slot on September 22, 2026 at a lower $4 in / $20 out per MTok.

Data verified September 23, 2026 against Anthropic's launch announcement. Anthropic did not publish a context-window figure at launch, so this page leaves it unstated rather than guessing, freshness and honesty are the whole point of these model pages.

Anthropic's launch numbers are framed as relative claims rather than raw absolute scores, so that is how we report them, attributed and unrounded. Every figure below is Anthropic's own, from the Opus 5 launch materials:

Data verified September 23, 2026. These were launch-day vendor claims, so treat them as provisional, as with any vendor benchmark. Two of the four (CursorBench and OSWorld) are explicitly framed against Fable 5, which is the more interesting comparison and the one we break down next.

This was the easy call at launch. Opus 5 and Opus 4.8 cost the exact same $5 in / $25 out per MTok, and Anthropic reported Opus 5 more than doubling Opus 4.8's Frontier-Bench v0.1 score at a lower cost per task. For new work today, skip both: Opus 5.5 costs less than either at $4 in / $20 out per MTok.

Fable 5 was Anthropic's Mythos-class frontier model and still topped some benchmarks, but it cost $10 in / $50 out per MTok, twice Opus 5's $5 in / $25 out per MTok. Opus 5's pitch was that you gave up very little to save half:

The read at the time: for the large majority of agentic coding, Opus 5 was the value pick, near the frontier at half the price, with Fable 5 kept for the genuinely hardest work. The same split holds one generation on, with Opus 5.5 as the default and Fable 5.1 ($10 in / $50 out per MTok) as the escalation; the current rule is in the best Claude model for coding. The archived breakdown, including cost math, is in Claude Opus 5 vs Claude Fable 5.

Opus 5 was the target of /model opus until Claude Code v2.1.280, which pointed the alias at Opus 5.5. To keep a session on Opus 5, select it by its full model ID:

/model claude-opus-5

Opus 5's Fast mode ran at around 2.5x the default speed for twice the base price, through usage credits. If you run agents in parallel, routing is where the Opus tier earns its keep: on an AIDEN board each feature card launches its own Claude Code session, so you can hand Opus 5.5 the complex features, or a pinned Opus 5 where a workload was validated on it, drop to Sonnet 5.5 or Haiku 5.5 for mechanical ones, and reserve Fable 5.1 for the one genuinely brutal refactor. That per-feature routing is the whole point of model routing for coding agents.