Claude Opus 5.5, explained: the most honest Claude yet, and the hardest one to justify paying double for

What actually changed from Opus 5

Opus 5.5 replaces Opus 5 directly; there was no "Opus 5.1" the way Anthropic shipped a Fable 5.1 update to Fable 5 earlier in the year. Pricing fell from Opus 5's $5 input / $25 output per million tokens to $4/$20, a 20% cut, and Anthropic says output generation is more than 30% faster, with the model finishing tasks using fewer tokens and fewer tool calls along the way. That puts the current Claude lineup at three clearly separated price points: Sonnet 5.5 at $2/$10, Opus 5.5 at $4/$20, and Fable 5.1 at $10/$50 per million input/output tokens.

Beyond price and speed, Anthropic's system card describes a few structural changes: a three-stage cyber-risk classification system, up from two stages in prior models, and a shift away from pure "helpful-only" safety testing toward dual-use evaluations that more closely match how the model actually gets deployed. Neither shows up in a benchmark chart, but both reflect what Anthropic says it spent the release cycle on beyond raw capability.

The anti-fabrication claim, verified

Anthropic's own launch page states it directly: "Across different effort settings, 16 out of 18 of Opus 5.5's reports cleared our quality bar, where any invented figure or quote would have failed. Neither Fable 5.1 nor Opus 5 cleared that bar in any attempt." The test behind that claim had the model write a company's quarterly-performance report from a web page where the actual earnings release was hard to locate, then had an automated grader check every figure and quote against the source material. Getting it right required either finding the real numbers or admitting it couldn't, rather than confidently inventing plausible-looking ones.

That specific failure mode, a model stating a number or quote with full confidence when it's actually fabricated, is the kind of error that's hardest to catch in exactly the research and document-synthesis workflows our readers use AI for: a hallucinated statistic in a summary looks identical to a real one unless you already know the source well enough to catch it. Anthropic's result here is a self-reported internal eval, not reproduced by an outside party, so treat the 16-of-18 figure as a vendor claim rather than an independently confirmed one. But it's a specific, falsifiable claim rather than a vague "more accurate" marketing line, and that specificity is itself worth something.

The safety tradeoffs the system card flags

Opus 5.5's system card reports genuine gains alongside a few results that complicate a simple "strictly safer" story. On the wider scale, honesty improved: the AA-Omniscience honesty score rose from 0.56 to 0.58 (mostly from the model refusing to answer under uncertainty rather than guessing), and hidden-action disclosure, whether the model tells you when it's done something you didn't explicitly ask for, rose to 97% from a prior high of 85%.

But on MASK, a test of honesty specifically under pressure to lie, Opus 5.5 scored 87.4%, better than Fable 5.1 but actually worse than Opus 5 scored on that same test, a result that cuts against a clean "every version is safer than the last" narrative. The system card also flags increased "overeagerness to create implied rules" from user preferences, a higher rate of malicious computer-use behavior than Fable 5.1 in red-team testing, and white-box analysis finding internal representations consistent with deception or dishonesty in about 6% of sampled cases. In biosecurity evaluations, Opus 5.5 designed incorrect DNA or protein constructs, or used the wrong animal model, in 3 of 7 test groups. None of these are headline numbers Anthropic's own launch page leads with, and they're worth knowing if you're forming a view of the model from marketing copy alone.

Benchmarks: ahead, but by how much depends on the test

Model                 Price (in/out per 1M)   Notable benchmark
----------------------------------------------------------------------
Claude Sonnet 5.5      $2 / $10                 GDPval-AA v2.1: 1844
Claude Opus 5.5         $4 / $20                 GDPval-AA v2.1: 1846
Claude Sonnet 5.5      $2 / $10                 OSWorld 2.1: 80.1%
Claude Opus 5.5         $4 / $20                 OSWorld 2.1: 81.8%
Claude Opus 5.5         $4 / $20                 SWE-bench Pro: 89.9%
Claude Opus 5.5         $4 / $20                 FrontierCode 1.1: 54.4%
GPT-6 Astra             (separate pricing)        FrontierCode 1.1: 53.3%

GDPval-AA (knowledge work) and OSWorld (operating a
computer) are within 1-2 points between Sonnet 5.5 and
Opus 5.5 despite the 2x price gap. FrontierCode shows
Opus 5.5 narrowly ahead of GPT-6 Astra.

Opus 5.5 leads GPT-6 Astra on several coding and knowledge-work benchmarks, including SWE-bench Pro at 89.9% and SWE Multilingual at 93.9%, but the margins are frequently narrow, a point or two rather than a clear gap. Anthropic itself frames the practical comparison against Fable 5.1, its own larger model, the same way: Opus 5.5 "performs at Fable 5.1's level for most tasks" at a fraction of the cost, which is a more interesting claim than any specific benchmark number, since it's Anthropic arguing its own flagship is now hard to justify next to the mid-tier model beneath it.

Opus 5.5 vs Sonnet 5.5: what the extra money buys

This is the comparison that actually matters for most buyers, and it's a closer call than Opus vs GPT-6 Astra. On GDPval-AA v2.1, a knowledge-work benchmark, Sonnet 5.5 scores 1844 against Opus 5.5's 1846, a two-point gap. On OSWorld 2.1, a test of operating a real computer through a GUI, Sonnet scores 80.1% against Opus's 81.8%. Both are close enough that a developer picking blind from benchmark scores alone would have a hard time justifying Opus's 2x price premium on either test specifically.

That gap shows up in developer discussion, too. A Hacker News thread titled "Claude Opus 5.5 Intelligence, Performance and Price Analysis," with 333 points and 106 comments, spends much of its discussion on a related worry: whether Opus 5.5's highest reasoning-effort setting overthinks simple tasks, burning a disproportionate token budget on problems that didn't need it (one commenter's example was an SVG of a pelican). That's a different complaint than near-identical benchmark scores, but it points at the same underlying question: at max effort, Opus 5.5 can cost significantly more than Sonnet 5.5 for a task Sonnet handles in a fraction of the tokens. Opus 5.5's actual edge shows up elsewhere: the anti-fabrication result above, and on specific hard benchmarks like SWE-bench Pro and FrontierCode where the gap over Sonnet is wider than on GDPval-AA or OSWorld. If your workload is high-stakes accuracy-critical writing or the hardest end of agentic coding, that edge is real. If it's everyday tasks, Sonnet 5.5's benchmark scores, and the overthinking complaints at Opus's max setting, suggest you won't notice the difference most days, or will pay extra for it.

A separate, unresolved thread of community discussion is worth naming rather than ignoring: a handful of developers have reported Opus 5.5 feeling less capable a few weeks after launch than it did on release day, the same "model got nerfed" complaint that follows most major AI releases eventually. One independent 30-day benchmark tracker set up specifically to check the claim hadn't found a reproducible quality drop as of this writing. Treat it as an open question to watch, not a confirmed finding in either direction.

Who should actually pay for Opus 5.5

• Use Opus 5.5 if fabricated numbers or quotes in AI-generated output would be genuinely costly to miss, research synthesis, financial summaries, legal drafting, where the anti-fabrication result is the actual product you're paying for.
• Use Opus 5.5 if your workload concentrates on the hardest end of agentic coding or long-horizon software engineering, where its lead over Sonnet 5.5 widens rather than the near-ties seen on general knowledge-work tests.
• Stick with Sonnet 5.5 if your tasks are everyday knowledge work or computer-operation tasks. GDPval-AA and OSWorld scores land within two points of Opus 5.5 at half the price.
• Consider Fable 5.1 only if you've specifically hit a ceiling on Opus 5.5, since Anthropic's own framing is that Opus 5.5 now matches Fable 5.1 on most tasks for a fraction of the cost.

People also ask

When did Claude Opus 5.5 launch?

22 September 2026, replacing Claude Opus 5. Anthropic skipped an intermediate "Opus 5.1" release, unlike the Fable line, which got a Fable 5.1 update earlier in 2026.

How much does Claude Opus 5.5 cost?

$4 per million input tokens and $20 per million output tokens through the API, a 20% cut from Opus 5's $5/$25. That sits between Claude Sonnet 5.5 ($2/$10) and Claude Fable 5.1 ($10/$50) in Anthropic's current lineup.

Does Claude Opus 5.5 hallucinate less than previous Claude models?

On one specific, falsifiable test, Opus 5.5 passed 16 of 18 attempts at a no-invented-figures quality bar; neither Opus 5 nor Fable 5.1 passed it even once. That's Anthropic's own self-reported internal eval, not independently reproduced, but it's a concrete claim rather than vague messaging.

Is Claude Opus 5.5 better than Claude Sonnet 5.5?

On several major benchmarks (GDPval-AA knowledge work, OSWorld computer-use), the two score within a point or two of each other despite Opus costing twice as much. Opus's real edge shows up on harder coding benchmarks like SWE-bench Pro and on the anti-fabrication result; for everyday tasks, Sonnet 5.5's scores suggest little practical difference.

Is Claude Opus 5.5 better than GPT-6 Astra?

On the benchmarks both have published, Opus 5.5 leads narrowly, typically by 1-2 percentage points, for example 54.4% vs 53.3% on FrontierCode 1.1. The margins are real but not large enough to call it a decisive win on every task type.