Opus 5 was meant to stay behind Fable 5. It beats it on 8 evaluations out of 13

Claude Opus 5 official benchmarks, evaluation by evaluation

Below is the full set of evaluations published by Anthropic at the launch of Opus 5, against the three comparison models used in the official table: Fable 5 (the flagship released three weeks earlier), Opus 4.8 (the direct predecessor) and GPT-5.6 Sol (the most-cited competitor). Cells where Opus 5 does not come first are flagged — they are more instructive than the rest.

Source: official Anthropic benchmark table, published at the launch of Claude Opus 5 (24/07/2026). The health and biology rows compare against Mythos 5 where the official table does so. A dash marks an evaluation not reported for that model.

How much does Claude Opus 5 cost?

$5 per million input tokens, $25 per million output tokens — precisely the Opus 4.8 rate. A fast mode, roughly 2.5× quicker, is billed at double. The model is available through the API under the identifier claude-opus-5, becomes the default model on the Claude Max subscription and the most capable one available on Claude Pro. Standard access carries no data-retention requirement (Clubic and BelieveMy, 24/07/2026).

So the sticker price is unchanged. The cost per task, however, collapses — and that is where the news actually is. The CursorBench results published by BridgeMind on 24 July put exact numbers on it, and they say more than any isolated score.

Source: CursorBench results published by BridgeMind (@bridgemindai), 24/07/2026.

Three readings stand out. One: Opus 5 Max finishes half a point behind Fable 5 Max — 70.0% against 70.5% — at less than half the cost per task ($8.23 against $17.32) and roughly 40% fewer tokens. Two: drop one effort level and Opus 5 Extra High beats Fable 5 Extra High (69.3% against 68.4%) while costing $7.35 against $11.73. Three, the most interesting line for a tight budget: Opus 5 High scores 66.7% for $3.91 per task, half a point behind GPT-5.6 Sol Max which costs $5.69.

This is where the effort parameter stops being a technical detail: the same model family spans $3.91 to $8.23 per task for three to four points of score. On a workflow running thousands of times a month, that slider is the architecture decision. The same mechanism shows on OSWorld 2.0, where Opus 5 beats Fable 5's best score at a little over a third of the cost.

On safety, Anthropic claims its most aligned model to date: the rate of misaligned behaviour is the lowest measured across its recent models (a score of 2.3 in the alignment audit), and safety classifiers trigger roughly 85% less often than on Fable 5 — a practical point for anyone who abandoned a deployment because of false positives.

Opus 5 vs Fable 5 vs GPT-5.6 Sol: where are the real gaps?

The lazy reading counts the winning cells and concludes "new best model". The useful reading looks at where the gaps are large, and where they are not. Sorted by the size of the jump between Opus 4.8 and Opus 5, the table tells a clear story.

Every jump is on the same side: autonomy. Novel problem-solving (ARC-AGI-3): from 1.5% to 30.2%. Agentic terminal coding (Frontier-Bench): from 21.1% to 43.3%, a doubling. Computer use (OSWorld 2.0): from 55.7% to 70.6%. Business workflows (AutomationBench): from 17.0% to 26.0%. These four evaluations share one property — they measure the ability to chain actions in a real environment without a human taking over at every step.

Every near-tie is on the same side too: "classic" coding and expert domains. On DeepSWE v1.1, GPT-5.6 Sol stays ahead (72.7% vs 68.8%). On FrontierCode v1.1, Fable 5 edges out Opus 5 by a tenth of a point (53.5% vs 53.4%) — not a real gap, but enough to forbid writing that Opus 5 "dominates coding". On legal work, Fable 5 does better (13.3% vs 11.7%); on health, Mythos 5 dominates (66.0% vs 59.8%), and even GPT-5.6 Sol comes ahead of Opus 5 (60.5%).

This relative plateau on pure coding deserves to be stated plainly, because it contradicts the prevailing narrative: across Fable 5, Opus 5 and GPT-5.6 Sol, the spread on classic coding evaluations is about five points. Picking a model will not determine the quality of a custom software development project — architecture, tests and domain knowledge will continue to do that.

Auditing our four bets — including the one we lost

On 15 July we published four signed, dated bets on Opus 5, promising they could be held against us. The model has shipped: here is the outcome, including where it is unflattering.

Bet 1 — "Opus 5 will ship before the end of July": won

Written on 15 July, when prediction markets gave it roughly even odds and no primary source existed. The model shipped on 23 July. Nothing heroic: it was the most probable inference from Fable 5's commercial calendar.

Bet 2 — "it will be cheaper than Opus 4.8": lost on the letter, won on the substance

Lost in the strict sense: the rate card is identical — $5 / $25 per million tokens, exactly Opus 4.8's pricing. We bet on a lower sticker price; that did not happen.

But the economic intuition was right, and that is what makes this bet worth re-reading. We wrote that "the battle is fought on cost per task, not on benchmark records". That is precisely what Anthropic shipped: at an unchanged sticker price, Opus 5 reaches Fable 5's level at half the cost per task on CursorBench 3.2, and beats it on OSWorld 2.0 at a little over a third. The price cut did happen — not on the label, on the bill.

Bet 3 — "a fourth extension for Fable 5": half right

Anthropic did not extend free access: on 18 July it announced something else — Fable 5 stays in Max and Team Premium, but at reduced limits, while Pro plans lose included access. We had the direction (Anthropic is not cutting Fable 5), not the shape.

Bet 4 — "GPT-6 has nothing to do with this": won

The official table compares Opus 5 with GPT-5.6 Sol. No GPT-6 anywhere, one week after half the AI press invoked it as the driving force.

How good is Opus 5 at business workflows? 26% on AutomationBench

This is the row nobody will discuss, and the only one that describes our trade. AutomationBench measures business workflows: chains of enterprise tasks, not laboratory puzzles. Opus 5 scores 26.0% there, against 17.4% for Fable 5, 18.1% for GPT-5.6 Sol and 17.0% for Opus 4.8.

Both halves of that number deserve a sentence.

The good half: the gap to the pack is the widest in the entire table. Where other evaluations separate models by a few points, this one separates them by half. If you automate processes, model choice has just stopped being neutral.

The bad half, and it matters more: 26% is roughly one task in four. The best model in the world, on the evaluation that most resembles a real enterprise process, fails three business workflows out of four. Every commercial promise along the lines of "AI automates your company" runs into this number, published by the vendor itself.

The figure that really decides a production rollout: prompt injection

This is the point Boris Cherny, creator of Claude Code at Anthropic, puts ahead of every evaluation score — and he is right, because it is the one that decides whether you can point an agent at your real data. In his words, Opus 5 is the hardest model Anthropic has produced to hijack through prompt injection, a result he notes is somewhat buried in the system card.

What are we talking about? Indirect prompt injection is the attack where content your agent reads — an email, a supplier PDF, a web page, a product sheet — carries hidden instructions that hijack its behaviour. It is the number-one risk for any agent processing inbound documents, and the reason many automation projects never leave the pilot stage.

The Gray Swan IPI benchmark measures exactly that: the probability an attack succeeds within k attempts. Lower is better.

Source: Gray Swan IPI benchmark, chart published by Boris Cherny (Anthropic), 24/07/2026. Read it as: the lower the percentage, the more resistant the model.

The gap here is of a different order than on any other evaluation. On a single attempt, Opus 5 drops to 0.2% against 3.1% for GPT-5.6 Sol and 14.2% for Gemini 3.1 Pro. More importantly, it holds up over time: after fifteen attempts the success probability stays at 2.0%, where it climbs to 20% for GPT-5.6 Sol and past 60% for Grok 4.5 and Gemini 3.5 Flash. Cherny adds that layering defences — model alignment, injection probes and automatic mode — drives the attack success rate to roughly zero.

The result drew praise beyond Anthropic: Elon Musk replied "very impressive" to Cherny's post within minutes. The remark is worth noting, because his own company's model, Grok 4.5, sits at the bottom of this ranking — a 13.4% attack success rate on the first attempt, 60.8% after fifteen. When a direct competitor publicly endorses a figure that works against them, the figure is usually solid.

What early users are saying: "a hard model to love"

It would be dishonest to relay only the enthusiasm. The two most serious independent reviews on launch day are critical — and far more useful than the benchmarks for anyone deciding on a migration.

Claire Vo ran Opus 5 through her own benchmark: seven tasks, seven competing models including GPT-5.6 Sol, Sonnet 5 and Gemini 3.1 Pro, blind scoring. Her verdict is in her title — brilliant, but annoying. Opus 5 earns top marks on at least one use case while proving verbose and "neurotic", to the point of refusing to touch a merge conflict during a real coding session, out of excessive caution.

Dan Shipper (Every) went further: a week of testing across coding, writing, knowledge work and an internal agent. His starting assessment is harsh — the model argues with instructions, stops before the work is finished, and does not play well with existing tooling and automations. Then his team did something counter-intuitive: they deleted everything and started from scratch. Without the elaborate workflows built for earlier generations, Opus 5 became "dramatically better".

Shipper closes with a barbed image: Opus 5 has "the personality of the genius" without its top end, leaving it in an uncomfortable middle ground between the elite model and the fast generalist.

The most rigorous bench: CodeRabbit

A third source, and the most methodical: CodeRabbit ran Opus 5 through its code-review bench — around a hundred error patterns drawn from real open-source pull requests, three runs per configuration, compared against their production model mix. They have benched every Opus release since version 4, which gives rare perspective.

Their verdict is nuanced and worth quoting as-is: a precision specialist, to be paired with a recall model. The numbers:

Source: CodeRabbit, "Opus 5 model review", 24/07/2026 — ~100 error patterns from real pull requests, averages across three runs.

In other words: Opus 5 says fewer wrong things when it speaks, but it misses more bugs and produces four times as many minor remarks to triage. CodeRabbit draws an architectural conclusion from this, not a ranking: frontier models have split into two camps — recall-oriented ones that find a lot and lean on filtering (GPT-5.6 Sol catches 69.7% of bugs, but only 31.6% of its comments are worth keeping), and precision-oriented ones that say less and are right more often. Opus 5 anchors the second camp.

Two figures matter for a budget. First, Opus 5 reads about 50% more and writes about 65% more than the baseline models for the same job (≈60,500 input and 9,500 output tokens per call, against ≈40,500 and 5,800): part of the real premium comes from volume, not from the rate card. Second, the context window moves to one million tokens by default, with performance claimed to hold throughout — a direct answer to the degradation past 200,000 tokens observed on Opus 4.8.

By category, CodeRabbit finds it solid on configuration errors and code quality, weaker on logic errors, race conditions and API misuse. Practical translation: good as an additional perspective, insufficient as the only safety net on critical code.

These criticisms match our own reading from the field: a more autonomous, more cautious model is not automatically a more useful one in production. An agent that refuses to act when in doubt beats an agent that breaks something — but every refusal is an exception to handle by hand, and therefore one more line in the 74% discussed above.

What Opus 5 actually changes for a mid-sized company

Three operational consequences, without extrapolation.

1. Autonomous agent projects become defensible — some of them, not all. An agent that must chain steps across your tools benefits directly from the jumps on OSWorld 2.0 and Frontier-Bench. A project shelved six months ago for "too many steps, too many manual recoveries" deserves a second look. A project shelved because the process itself was never written down remains exactly as bad.

2. Model choice becomes an architecture decision, not a preference. The table shows domain-by-domain inversions: legal goes to Fable 5, health to Mythos 5, pure coding to GPT-5.6 Sol on DeepSWE. A serious architecture routes each task to the model that fits, instead of locking in a single vendor — which is what we do in our AI automation projects, where the intelligence layer is explicitly interchangeable. It is also what makes a switch to an open-weight model practical the day it becomes relevant — Kimi K3 being the current case in point.

3. None of this changes your obligations. The European AI Act deadlines stay on the same dates: transparency (Article 50) on 2 August 2026, high-risk Annex III systems on 2 December 2027. A more autonomous model mechanically means more decisions taken without a human in the loop — so more traceability to plan for, not less.

Beyond that, the method does not move: map the tasks, quantify the return on investment, start with the process whose failure costs least, and put nothing into production without a rollback plan. The 26% on AutomationBench is the best reason to stick to it.

FAQ — Claude Opus 5