Claude Opus 5 Benchmarks: Where It Wins and Loses

Claude Opus 5 Benchmarks: Where It Wins and Loses

The first thing I did when Claude Opus 5 landed was not read the benchmark chart.

Short version, for anyone who just wants the answer: the Claude Opus 5 benchmarks show it beating Fable 5 on agentic and long-running work at half the token price — and losing to it on five evals nobody is citing. It costs $5/$25 per million tokens, same as Opus 4.8.

But before any of that is usable, there's a gate. I opened a terminal and ran claude --version. It printed 2.1.219. I exhaled, because that number is the entire difference between reading about this model and using it — Opus 5 requires Claude Code v2.1.219 or later, and anything older doesn't just perform worse. It doesn't show the model at all. No row in the picker. No error explaining why. Just a list that quietly stops at Opus 4.8, and a developer wondering whether Anthropic forgot about their account.

I got lucky. My install had auto-updated the night before. If yours hasn't, every Claude Opus 5 benchmark in this article is theoretical for you until you run claude update.

That's a strange thing to lead with on a model launch. But it's the first of several things about this release that the coverage has skipped, and I've spent the day since launch working through the rest of them — the five benchmarks where Opus 5 loses, the Opus 4.8 score that three different outlets report three different ways, the effort setting that behaves differently on this model than on any Claude before it, and the fallback routing change that quietly affects how you should read the published numbers at all.

Anthropic's framing is "nearly all of Fable 5's intelligence at half the price." That framing is mostly true. It is also doing a lot of work.

What Actually Shipped on July 24

Claude Opus 5 went live on Friday, July 24, 2026, as claude-opus-5 — a fixed model ID, no date suffix, same scheme as claude-opus-4-8. It landed simultaneously on the Claude API, Amazon Bedrock (anthropic.claude-opus-5), Google Cloud, and Microsoft Foundry, plus Claude.ai, Claude Code, and Claude Cowork.

The specs, before we get to anything contested:

Two rows there deserve more attention than they've gotten.

The price didn't move. Opus 5 costs exactly what Opus 4.8 cost. For a model Anthropic is positioning as a step change, that's the second consecutive Opus release where the per-token sticker held flat. If you've been running 4.8 as a daily driver, your bill does not change on switching — which makes the upgrade decision unusually clean, because the only variable left is token volume, not token price.

The prompt-cache minimum halved. Opus 5 caches prompts as short as 512 tokens; Opus 4.8 needed 1024. This sounds like trivia until you run agent loops. Every system prompt, tool definition block, and repeated preamble that sat in the 512–1023 token dead zone on 4.8 — too short to cache, so re-billed at full input price on every single turn — now caches. On a tight agent loop firing hundreds of turns, that's a real line-item change nobody put in a headline. It's the same class of saving I chased manually when I wrote up five ways to cut Claude Fable 5 usage costs, except this one arrives for free.

Now the part everyone leads with.

The Claude Opus 5 Benchmarks — Full Table

Here are the published figures, sourced from Anthropic's announcement and the independent coverage that reproduced them. I'm giving you the whole board rather than the flattering slice, because the losses are where the useful information is.

* See the next section. This one number is reported differently depending on where you read it.

Two results genuinely deserve the attention they're getting.

Frontier-Bench v0.1 is agentic terminal coding — the closest published proxy for what Claude Code actually does all day. Opus 5 scores 43.3%, against 33.7% for Fable 5 and 34.4% for GPT-5.6 Sol. Beating a model that costs twice as much on the benchmark closest to your primary workload is the strongest single argument in this launch, and it's the one I find most credible because the task shape matches the use case.

ARC-AGI-3 is the eye-catcher: 30.2%, against 1.5% for Opus 4.8. That's a twentyfold jump on an eval specifically built to resist memorization — it tests whether a model can infer novel rules it has never encountered. GPT-5.6 Sol manages 7.8%.

Here's the caveat nobody attaches to that number: Fable 5 wasn't tested on ARC-AGI-3. Anthropic's own most capable model has no published score on the benchmark being used to demonstrate the generational leap. The comparison is Opus 5 against its predecessor and against OpenAI — not against the model it's being priced beneath. That's not dishonest. It is an incomplete picture presented as a complete one, and if you're using this chart to choose between Opus 5 and Fable 5, that row tells you nothing.

Why Three Outlets Report Opus 4.8 Differently

Watch that asterisk in the table. On Frontier-Bench v0.1, Opus 4.8's score is reported as:

• 18.7% by VentureBeat and Vellum
• 21.1% by The Decoder, Decrypt, and most aggregators

That's a 2.4-point spread on the same model, on the same benchmark, in coverage published within hours of each other. Nobody I've read flags it.

The most likely explanation is effort levels. Anthropic's benchmark tables report scores at specific effort settings, and Opus 4.8 supports five — low through max. A model evaluated at high and the same model evaluated at xhigh are two different data points wearing one name. If 18.7% is Opus 4.8 at its default and 21.1% is Opus 4.8 pushed harder, both numbers are correct and neither is comparable to the other without the setting attached.

I can't confirm which reading is right, and I'm not going to pretend otherwise. What I can tell you is the practical consequence: every headline comparison in this launch is a single point plucked off a curve. "More than double Opus 4.8" is 43.3 vs 18.7. "Roughly double" is 43.3 vs 21.1. Same model, same launch, different story, depending on which row of the effort ladder someone screenshotted.

Which brings us to the thing the marketing genuinely obscures.

"Half the Price" Is a Claim About Effort, Not Tokens

Read the pricing table again. Opus 5 is $5/$25. Fable 5 is $10/$50. So Opus 5 is literally half of Fable 5 per token, and the headline is accurate on its face.

But per-token price is not what you pay. What you pay is per-token price multiplied by tokens consumed, and tokens consumed is governed by the effort level — which is exactly where the comparison gets slippery.

Anthropic's own framing on CursorBench 3.2 is that Opus 5 lands within 0.5% of Fable 5 at max effort, at half the cost per task. Parse that carefully. It isn't "Opus 5 matches Fable 5 at identical settings for half the money." It's that Opus 5 reaches competitive results at a lower point on its own effort curve — so the saving comes from needing less compute to get there, not from a straight per-token discount at parity.

That distinction matters the moment you set effort yourself. Run Opus 5 at max on everything and the token multiplier can erase the sticker advantage entirely. Run it at medium where quality holds and the saving compounds. It's the same lesson I hit with Grok 4.5 as a cheap coding daily driver — the number that decides your bill is token count, not the price per token. The Anthropic migration guidance is unusually direct about this: low and medium on Opus 5 are strong enough that prior-model effort defaults rarely transfer, and a fresh sweep is the single highest-leverage tuning step after switching.

Two other numbers in the chart deserve the same skepticism, and this is the part I think is genuinely underreported.

Look at the absolute values, not the rankings. The legal agent benchmark tops out at 13.3% — for the winner. Business workflows on AutomationBench: 26.0% for the leader. These are the domains being marketed as knowledge-work breakthroughs, and every model tested is failing roughly three-quarters to nine-tenths of the tasks. Opus 5 ranking first at 26% is a real result. It is not a result that means "you can hand your business workflows to an agent." Ranking language hides how much runway is left.

And there's a contamination question on the safety fallbacks. Both Opus 5 and Fable 5 run safety classifiers that can decline a request, and both route flagged requests to Opus 4.8 as a fallback. Which means published scores on any benchmark touching security or biology content may include runs where the model under test silently handed off to an older model. I have not seen Anthropic disclose how flagged-and-rerouted requests are accounted for in the published figures. That's a legitimate open question about the methodology, not an accusation — but if you're evaluating for security work specifically, it's the question I'd want answered.

The Claude Opus 5 Benchmarks Nobody Screenshots

Every launch post I read led with Frontier-Bench and ARC-AGI-3. Here's the other half of the board.

1. DeepSWE v1.1 — agentic coding. Opus 5: 68.8%. Fable 5: 69.7%. GPT-5.6 Sol: 72.7%. Opus 5 finishes third on an agentic coding benchmark, roughly four points behind OpenAI's model and a point behind Fable 5. Given that agentic coding is the headline use case, this is the most important row in the entire table and it appears in almost no coverage.

2. FrontierCode v1.1. Opus 5: 53.4%. Fable 5: 53.5%. Statistically a tie, but it's a tie — not the blowout the Frontier-Bench framing implies.

3. Humanity's Last Exam, no tools. Opus 5: 56.3%. Fable 5: 56.5%. Another photo finish that goes the other way.

4. Legal agent benchmark. Opus 5: 11.7%. Fable 5: 13.3%. Fable keeps the edge in legal reasoning.

5. HealthBench Professional. Opus 5: 59.8%. Mythos 5: 66.0%. A six-point gap in clinical and health-professional tasks.

Stack those against the wins and a clean pattern falls out. Opus 5 dominates on long-running agentic work, computer use, business workflow execution, and novel problem-solving — tasks that unfold over many steps and require interacting with an environment. It ties or loses on single-shot, deep-domain reasoning — law, medicine, hard exam questions, and one-pass coding evaluations.

That is a genuinely useful mental model, and it's more actionable than any headline. Opus 5 is not "Fable 5 for half price." Opus 5 is a model tuned for duration and iteration, where Fable 5 is tuned for depth on a single hard thing. Pick by task shape, not by chart position. If your work is overnight refactors and multi-file agent runs, Opus 5 is the better instrument even where Fable scores higher. If your work is one dense legal or clinical judgment call, the chart is telling you to stay on Fable — a nuance that also showed up when I looked at how GPT-5.6, Grok 4.5, and Fable 5 split on creative work.

The Verification Loop — And Why You Should Delete Your Verification Prompts

The capability everyone is pointing at is self-verification. The example Anthropic published is the one that circulated: asked to rebuild a machine part as 3D code without being allowed to view the drawing, Opus 5 wrote its own computer vision pipeline to extract the geometry from raw pixels, then reconstructed the part from what it had extracted.

That's the behavior the agentic benchmarks are measuring — a model that builds its own checking apparatus mid-task rather than producing output and stopping.

Here's the counterintuitive part, and it's the single most practical thing in this article.

Because Opus 5 verifies its own work unprompted, telling it to verify makes things worse. Anthropic's own migration guidance says it plainly: instructions like "include a final verification step" or "use a subagent to verify" now cause over-verification on this model, and removing them reduces the problem with no capability regression. Per-prompt phrasing has the same effect — "double-check your answer," "re-verify before responding."

Sit with that for a second, because it inverts standard prompting advice. "Ask the model to self-check" has been sound guidance for two years. On Opus 5 it is actively counterproductive. If you maintain a prompt library, a CLAUDE.md, or a harness with a separate verification pass, that scaffolding is now overhead — it's paying tokens for work the model already does, and pushing it to redo work it already completed.

The migration action is a delete, not a rewrite. Go find your verification instructions and remove them.

A few related behavioral shifts worth knowing before you switch:

• It writes longer. Both conversational responses and files written to disk run longer than on 4.8. Lowering effort does not reliably shorten visible output — that's a prompting fix, not a settings fix.
• It expands scope. Opus 5 will sometimes add steps you didn't ask for. A short scope-discipline instruction handles it.
• It delegates more. This one reverses direction from 4.8, which under-reached for subagents and needed prompting to delegate. Opus 5 reaches for them freely — so if you added "delegate more" guidance for 4.8, take it back out and consider capping spawn counts, because each subagent re-establishes context and multiplies cost. If you've built out an agent-team setup, this is worth an audit before you switch models.

How to Actually Switch — Claude Code and the VS Code Extension

This is the section I wish had existed when I started, so here it is with the version gates attached.

Step 1 — check your version, because this blocks everything else.

claude --version

Mine returned 2.1.219 (Claude Code). That is the minimum build that can select Opus 5. If yours is lower, the model does not appear in the picker and no amount of typing will summon it.

claude update

For reference on the other version gates: Sonnet 5 needs v2.1.197 or later, Opus 4.8 needs v2.1.154. If you're several versions behind, updating unlocks all of them at once.

Step 2 — select the model. Three ways, in order of how sticky they are:

/model                    # opens the picker
/model opus               # alias — resolves to Opus 5 on the Anthropic API as of v2.1.219
/model claude-opus-5      # pins the exact version

The opus alias is worth understanding. As of v2.1.219 it resolves to Opus 5 on the Anthropic API and on Claude Platform on AWS, Amazon Bedrock, and Google Cloud's Agent Platform. Microsoft Foundry is the exception — opus still resolves to Opus 4.6 there, so Foundry users must name the model explicitly or set ANTHROPIC_DEFAULT_OPUS_MODEL. Before v2.1.219, opus meant Opus 4.8 on the Anthropic API, which is why an unattended claude update can silently change which model your team is running.

For a whole project, put it in settings instead:

{
  "model": "claude-opus-5"
}

Or launch a single session with it:

claude --model claude-opus-5

The VS Code extension is now generally available and supports the same /model and /context slash commands as the CLI, so the flow above works identically inside the editor. It's the same binary underneath — update the CLI and the extension follows.

Step 3 — and this is the trap — set your effort level explicitly.

Here is a genuine quirk, straight out of the Claude Code model-config docs, that I have not seen mentioned anywhere in the launch coverage:

When you first run Fable 5, Opus 4.8, or Opus 4.7, Claude Code applies that model's default effort even if you previously set a different level for another model, and holds it across sessions until you make an explicit effort choice. Opus 5 has no such hold: a level you previously set carries over.

Read that twice. Every other recent model resets you to its own sensible default on first use. Opus 5 does not. Whatever effort level you last set — for Sonnet 5, for Opus 4.6, for a one-off experiment you forgot about six weeks ago — silently carries over to Opus 5.

If you had low set for a latency-sensitive experiment, you are now benchmarking Opus 5 at low and wondering why the twentyfold ARC-AGI leap isn't showing up in your work.

Set it deliberately:

/effort              # opens an interactive slider
/effort xhigh        # recommended starting point for coding and agentic work
/effort auto         # resets to the model default (high)

The default is high on Opus 5. Anthropic's recommendation is to start at xhigh for coding and agentic work, then sweep downward — low and medium on this model are unusually strong, and the effort ladder is your primary cost lever, not the sticker price. At xhigh or max, give it room: max_tokens of at least 64K, because thinking and response text share that budget.

There's one more setting-level gotcha. On Opus 5, thinking is on by default — omitting the parameter runs adaptive thinking, where on Opus 4.8 omitting it meant no thinking at all. And thinking: {type: "disabled"} is only accepted at effort high or below; pair it with xhigh or max and the API returns a 400. If you're calling the API directly rather than through Claude Code and you have a route that ran thinking-off on 4.8 with a tightly-sized max_tokens, that route can now truncate mid-answer. Worth auditing before you flip the model string.

The Fallback Routing Change

One more thing that landed with v2.1.219, and it matters more than its footnote treatment suggests.

Opus 5 runs safety classifiers for cybersecurity and biology content. When one fires, Claude Code re-runs the request on a fallback model — but the routing is now category-specific, and it differs by which model refused:

• Opus 5: cybersecurity-flagged requests re-run on Opus 4.8. Biology-flagged requests end in a hard refusal — Opus 5 runs its own biology classifiers with no fallback target.
• Fable 5: biology-flagged requests re-run on Opus 5; cybersecurity-flagged requests re-run on Opus 4.8.

Before v2.1.219, every flagged Fable 5 request went to your provider's default Opus model, and Opus 5 wasn't a fallback source at all.

Two practical consequences. First, if you do security research — pentest tooling, CTF work, anything offensive-security adjacent — expect Opus 5 to hand off to Opus 4.8 frequently, sometimes on the very first request of a session, because that first request carries your CLAUDE.md and git status as context. A repo full of security material can trip the classifier before you type anything. claude --safe-mode disables customizations and tells you whether your own config is the trigger. This is expected routing for those domains, not an account problem — but it does mean the model answering you is often not the model you selected.

Second, you can take manual control: run /config and turn off "switch models when a message is flagged." A flagged request then pauses and asks whether to switch or let you edit the prompt and retry.

What I Haven't Verified Yet

I want to be exact about the line between what's established and what isn't, because the gap between "Anthropic published a number" and "I watched this work on my code" is the whole reason model reviews are worth reading.

Verified first-hand, today: the version gate. claude --version on my machine returns 2.1.219, which is the documented minimum for Opus 5. The /model and /effort mechanics, the alias resolution table, the effort-carryover behavior, and the fallback routing are all from Anthropic's current Claude Code model-configuration documentation, which I read rather than recalled.

Sourced, not run by me: every benchmark number above. Frontier-Bench, ARC-AGI-3, GDPval-AA, OSWorld, DeepSWE, and the rest are Anthropic's published figures plus independent reporting that reproduced them. I have reproduced none of them, and neither has anyone else who published within a day of launch. Treat the whole table as vendor-supplied until third-party replication lands — the same discipline I applied when Gemini 3.5 Pro arrived with leaks instead of a model card. A published number is a claim, not a result.

Genuinely open questions I'd want answered before drawing conclusions:

• Which effort level produces Opus 4.8's Frontier-Bench score, and therefore which of 18.7% or 21.1% is the like-for-like comparison?
• How are flagged-and-rerouted requests accounted for in published scores on security-adjacent benchmarks?
• Why does Fable 5 have no ARC-AGI-3 result, on the eval being used to demonstrate the generational leap?
• Does the DeepSWE third-place finish reproduce on real repository work, or is it a benchmark-harness artifact?

What I'm testing this week, in this order: the same Laravel service-extraction task I used to evaluate Opus 4.8's effort levels, run at medium and xhigh on Opus 5 to find where quality actually breaks down. Then a Next.js component migration with the verification scaffolding stripped out, to see whether "delete your verification prompts" holds up on real code. Then a FastAPI endpoint build with subagents capped, because the increased delegation is the change most likely to surprise my token bill.

I'll publish those numbers when I have them. Not before.

What I'd Actually Do Today

If you're running Opus 4.8 as your daily driver: switch. The price is identical, the agentic benchmarks favor Opus 5 substantially, and the prompt-cache change is free money on agent loops. Update to v2.1.219, run /model opus, then immediately run /effort and set it deliberately — because it did not reset for you, and a stale low from three weeks ago will make this model look like a downgrade.

If you're on Fable 5 for everything: split your workload rather than migrating it. Long-running agentic work, computer use, and multi-file builds go to Opus 5 at half the token cost. Dense single-shot reasoning in law, medicine, or hard analysis stays on Fable — the chart says so, in the rows nobody screenshotted. That split is easier to operate if your knowledge layer is already model-agnostic, which is most of the argument for building a Fable 5 second brain that isn't tied to one model.

If you're doing security research: understand that you are frequently going to be talking to Opus 4.8 whether you selected it or not, and plan around the handoff rather than being surprised by it.

And whatever you switch to, go delete your verification prompts. The instruction you added two years ago to make a weaker model check its work is now making a stronger one waste tokens rediscovering what it already knew.

That version number I checked before reading the benchmarks turned out to be the right instinct for the wrong reason. I thought I was checking whether I could run Opus 5. What I was actually looking at was the reason most people evaluating this model over the next month will draw the wrong conclusion from it — running an old build, or the right build at whatever effort level they happened to leave lying around, and deciding from that whether a twentyfold jump in novel problem-solving is real.

The benchmarks aren't the hard part. The setup is.

Let's Work Together

Looking to build AI systems, automate workflows, or scale your tech infrastructure? I'd love to help.

• Fiverr (custom builds & integrations): fiverr.com/s/EgxYmWD
• Portfolio: mejba.me
• Ramlit Limited (enterprise solutions): ramlit.com
• ColorPark (design & branding): colorpark.io
• xCyberSecurity (security services): xcybersecurity.io