Claude Opus 5: What Actually Changed — the Half-Price Model That Caught the Flagship

Claude Opus 5 arrived on July 24, 2026 with an unusual pitch: same price as its predecessor Claude Opus 4.8, and a claim to match Claude Fable 5 — Anthropic’s top model, which costs exactly twice as much — on many tasks. Claims like that deserve suspicion. So we went looking for independent measurements, and found both support and clear counterexamples. There is also one spec change that will silently truncate your responses if you ignore it.

Half price, matching the flagship: does it hold?

The prices first. Claude Opus 5 costs 5 dollars per million input tokens and 25 per million output — identical to Opus 4.8, no increase. Claude Fable 5 sits at 10 and 50, precisely double. On the independent side, the Artificial Analysis Intelligence Index reads as follows in early August 2026, at max effort.

On this one index, Opus 5 does not just match the flagship — it edges past it. The same firm reports Opus 5 ahead of Fable 5 on GDPval-AA v2 (1861 Elo, plus 114) and AA-Briefcase (1720 Elo, plus 146). But keep the nature of the yardstick in mind: a composite index blends math, science, coding, and reading tasks into one number, and reweighting the mix reshuffles the ranks. As we will see below, other measurement outfits reach the opposite conclusion. Arguably the more important fact in this table is that five models are packed between 63 and 57.

Where you can use it

Claude Opus 5 is available on every paid plan and via the Claude API, and it became the default model on Claude Max. Platforms cover the Claude API, Claude.ai (Pro, Max, Team, Enterprise), Amazon Bedrock, Google Cloud, and Microsoft Foundry, with GitHub Copilot support from day one. One caveat: the roughly 2.5x-speed fast mode (priced at 10 in and 50 out) is API-only and absent from Bedrock, Google Cloud, and Foundry.

Thinking is now on by default — the migration trap

Omit the thinking setting in a request and Opus 5 thinks; Opus 4.8 did the opposite. Swap the model name in existing code and you get thinking you did not ask for. That matters because the output cap covers thinking plus answer: limits tuned for no-thinking setups can cut the answer off mid-sentence, with no error raised. A second breaking change: explicitly disabling thinking is only accepted at effort high or below — combine it with xhigh or max and the request errors out. Audit both patterns before migrating.

Five effort levels, a 407-point spread

Opus 5 exposes five effort levels — low, medium, high, xhigh, max — defaulting to high. Artificial Analysis measured what the dial actually does: on GDPval-AA v2, effort alone moved the Elo score by 407 points, and output token volume spanned roughly 8x from low to max. Same model, same question, wildly different cost and quality. The official guidance: start at high, step up to xhigh for demanding coding and agent work, reserve max for when accuracy beats cost, and use low and medium liberally wherever quality holds. Interestingly, more is not always better — one analysis by MindStudio found a frontier coding benchmark plateauing around the 53 percent mark near medium, with higher effort costing more and scoring lower. A single analysis, but a useful caution against defaulting to max.

Cost per task: the inversion nobody advertises

Artificial Analysis also publishes what one Intelligence Index task actually costs to run. These figures are from the measurements right after launch in late July; since token prices are unchanged, the relative picture still holds.

Per-token prices are identical between Opus 5 and Opus 4.8, yet at max effort the new model costs more per completed task — it thinks longer, so it emits more tokens. The sticker price froze; your bill did not. The same firm notes, however, that at high and xhigh, Opus 5 can beat both Opus 4.8 and Sonnet 5 at a lower per-task cost. The efficient zone is one or two notches below the ceiling, which squares with the plateau finding above. Practical conclusion: switching to Opus 5 alone will not lower your bill — setting effort deliberately is what cashes in the frozen price.

Anthropic’s own numbers, read properly

By Anthropic’s own account, Opus 5 scores 43.3 percent on Frontier-Bench, 30.2 percent on ARC-AGI-3 (about triple the runner-up), 79.2 percent on SWE-bench Pro, and 70.57 percent on the computer-use benchmark OSWorld 2.0, up from 55.7 for Opus 4.8. That last jump is the interesting one — screen-driven workflows may have crossed a usability line. But vendor benchmarks are chosen terrain: read them as a statement of intended battlegrounds, not as comparisons. One number we deliberately omit: sources disagreed on the SWE-bench Verified score (96 versus about 97), so it does not appear here.

The numbers that did not make headlines

Now the inconvenient results. On AA-Omniscience, Opus 5 answers 7 points more accurately than Opus 4.8 — but its hallucination rate rose 14 points to 50 percent. It knows more, and admits ignorance less. Two mechanisms explain the paradox: a model that attempts harder questions has more chances to miss, and better prose makes wrong answers harder to doubt. For any fact-dependent work, the verification step just got more important, not less.

Opus 5 also loses on several yardsticks. Epoch AI’s Epoch Capabilities Index — built with item response theory, which resists ceiling effects — places it behind both Claude Fable 5 and GPT-5.6. The Vales Index, weighted by industry contribution to US GDP, has it slightly under Fable 5, and Cognition’s Apex SWE gives Fable 5 a narrow lead. So the honest summary is: Opus 5 beats the flagship on many measures, not all. For the flagship’s own design story, see our Claude summer catch-up.

Fewer refusals, and a landing pad

Anthropic says the safety classifiers fire about 85 percent less often than on Claude Fable 5, addressing complaints about legitimate security- and bioscience-adjacent work getting flagged. Alongside that comes Automatic Fallbacks, a beta feature that routes a flagged request to a less restricted model instead of erroring out. For production use the point is continuity: what used to require hand-written error handling is now a setting — the pattern we argued for in our fallback design piece.

Two quiet changes: cache and rate limits

The minimum cacheable prompt dropped from 1024 tokens on Opus 4.8 to 512. Prompts previously too short to cache may now qualify without code changes — worth rechecking if you resend fixed preambles. Separately, Opus 5 draws from a rate-limit pool distinct from the older Opus line, so confirm your tier’s ceiling before shifting production traffic.

Migration checklist from Opus 4.8

Find code that leaves thinking unset — it will now think, and tight output caps will truncate. Find spots that disable thinking while requesting xhigh or max effort — they will error. Assign effort per task, starting from high and pushing routine work down to low or medium; this is the step that actually lowers the bill. Raise output caps to cover thinking. Then verify rate limits and revisit cache settings for short prompts.

At the workbench

For 3D printing and design work, effort discipline is where the money is: changing one wall-thickness variable in a parametric enclosure does not deserve max, but re-deriving a clearance between a boss and a board does deserve xhigh — our hands-on CAD sessions show where that line runs. The million-token context rewards whole-project consultations: design notes, filament data sheets, failure logs, and slicer history in one pass, so answers come with your context instead of generic advice. The lower cache minimum helps if you prepend a fixed profile of your machines to every chat. And the raised hallucination rate is a direct risk for materials data — melting points, pilot hole sizes, standard dimensions — so keep verifying against manufacturer data sheets. The full workflow map is in our AI 3D design roadmap.

Verdict

Price frozen at half the flagship; first place on Artificial Analysis but trailing on Epoch and Apex SWE — so say it matched the top on many measures, and no more. Watch the thinking default and the effort-disable combination when migrating. Set effort per task, because the efficient zone is below max. And keep your fact-checking workflow, because a smarter model lies more convincingly. Of everything July changed, that last habit is the one that pays off first.

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】

swiftwand

X

AIを使って、毎日の生活をもっと快適にするアイデアや将来像を発信しています。 初心者にもわかりやすく、すぐに取り入れられる実践的な情報をお届けします。 Sharing ideas and visions for a better daily life with AI. Practical tips that anyone can start using right away.