Claude Opus 5 Matches Fable-Level Reasoning at Half the Token Price

Anthropic's Mid-Tier Bet: Near-Flagship Intelligence at Half the Price

When Anthropic released Claude Opus 5 on July 24, 2026, roughly two months after its predecessor Opus 4.8, the headline claim was unusually aggressive for a company that tends to underplay its mid-tier models: on most published benchmarks, the new Opus matches or beats the company's larger, pricier flagship Claude Fable 5 — while running at half the cost910. At $5 per million input tokens and $25 per million output tokens, Opus 5 priced identically to Opus 4.8 and exactly half of Fable 5's $10/$50 rate69.

That pricing structure matters because it inverts the usual assumption in model lineups — that the most expensive tier is meaningfully the most capable. Anthropic still positions Fable 5 as its most generally capable model, best suited for the hardest, longest autonomous tasks612. But the gap on paper has narrowed to the point where, for a large class of professional work, the cheaper model is simply the better buy.

The Reasoning Story: A Genuinely Large Leap

The most striking number in the Opus 5 launch was on ARC-AGI-3, a benchmark designed to resist saturation and test novel, abstract problem-solving. Opus 5 scored 30.2%, which Anthropic characterized as roughly three times the next-best model and nearly quadruple the previous frontier record — a jump that reads less like incremental tuning and more like a capability shift in how the model reasons about unfamiliar problems911. Fable 5 did not publish an ARC-AGI-3 score at all, making the comparison indirect but suggestive13.

On ARC-AGI-2, the two Anthropic models sit nearly level — 90.4% for Opus 5 versus 89.2% for Fable 5 — while on the original ARC-AGI-1, Fable 5 edges ahead 98.5% to 97.5%13. The pattern across the ARC family is consistent with what independent analysis found: Opus 5 made an unusually large leap in specific reasoning-heavy, novel-problem domains while remaining behind Fable-class models on broad aggregate capability indices12.

Multidisciplinary reasoning told a similar story. On Humanity's Last Exam, a benchmark built to defeat retrieval shortcuts with expert-level questions across subjects, Opus 5 scored 56.6% without tools and 63.6% with tools17. Fable 5 posted 57.8% and 63.8% on the same rows — a gap of roughly a point, well inside the standard error ranges Anthropic reported on comparable agentic benchmarks1718.

Where Opus 5 Actually Beat the Flagship

Several benchmarks flipped outright in the cheaper model's favor. On Frontier-Bench v0.1, Opus 5 scored 43.3% against Fable 5's 33.7%, and on SWE-bench Verified it led 96.0% to 95.0%11. On GDPval-AA v2, the Elo-style professional knowledge-work benchmark where Fable was supposed to have home-field advantage, Opus 5 posted 1,861 against Fable 5's 1,747 — a 114-point gap in favor of the model costing half as much11.

Computer use was another surprise. On OSWorld 2.0, Opus 5 surpassed Fable 5's best result at just over a third of the cost per task10. And on CursorBench 3.2, tested inside a real coding editor, Opus 5 at maximum effort landed within half a percentage point of Fable 5's peak while costing roughly half as much per task, beating every other model at a given cost across high, extra-high and max effort settings610.

Notably, Anthropic shipped Opus 5 without publishing SWE-bench Verified or Pro numbers of its own — the first time since 2024 its flagship coding model launched without one — leaving third parties to fill in the picture15.

The Reasoning Economics: Cost Per Task, Not Per Token

The deeper theme of the release is that per-token price is the wrong metric for reasoning models, because thinking tokens dominate the bill. Harvey, the legal-AI firm, reported that Opus 5 matched Opus 4.8's maximum-reasoning output quality while using 26% fewer tokens on average — the same answer, for meaningfully less money6.

Multiple independent testers found Opus 5 landed further left and higher on cost-versus-performance charts than Fable 5, GPT-5.6-Soul, and Opus 4.89. On the Apex Agents benchmark from Mercor, which grades frontier models on more than 200 real professional tasks like investment-banking analysis and management consulting against expert rubrics, Opus 5 led at 43.5%, marginally ahead of Fable 512.

The effort dial — a configurable reasoning-depth setting — turned out to be a genuine cost control rather than a gimmick. Later Anthropic analysis of the tier showed that up to about $2 per task a smaller model wins most comparisons, but at roughly $5 per task the Opus-tier model wins across the board, letting teams pay for depth only when a task demands it3.

Where the Coverage Diverges

The reporting does not fully agree on how dominant Opus 5 is. Anthropic's own framing and several launch analyses present it as matching or beating Fable 5 on most work911. Others note the picture is messier: Fable 5 retained a real edge on the very hardest, longest scientific reasoning, and on aggregate capability indices Opus 5 still trailed both Fable 5 and OpenAI's GPT-5.612. SWE-bench Pro was the one headline coding benchmark where Fable 5 stayed ahead, 80.3% to 79.2%11.

A cautionary note also emerged from the data: one comparison found Opus 5's higher overall score estimate versus Fable 5 (79.33 versus 78.82) had overlapping 90% confidence intervals — statistically not a decisive winner13. The honest read is that Opus 5 delivered value-tier dominance on work-shaped benchmarks, not blanket superiority.

The Sequel That Proved the Strategy

Two months later, the strategy Opus 5 pioneered became explicit. Fable 5.1 shipped in September and retook the lead on every category Anthropic published — but by margins of mostly one to four points, with one glaring exception: Terminal-Bench-Science 0.1, where Fable 5.1's 52.6% nearly doubled Opus 5's 29.0%1417. Then Opus 5.5 arrived on September 22 at $4/$20 per million tokens — cheaper than Opus 5 itself — and Anthropic described it as delivering Fable 5.1-level performance on most tasks at 40% less to run than Opus 52316.

Opus 5.5 beat Fable 5.1 on Terminal-Bench 4.0 (66.4% to 55.8%), GDPval-AA v2.1 (1,846 to 1,735 Elo), Humanity's Last Exam with tools (67.7% to 65.6%), and ARC-AGI-2, leading independent Artificial Analysis to rank it atop its intelligence index at 58 points against 53 for both Fable 5.1 and GPT-6 Astra41619.

That trajectory — each Opus generation closing on, then surpassing, the Fable tier it launched beneath — validates the thesis Opus 5 established in July: the mid-tier model is no longer a compromise, and the flagship premium is increasingly reserved for narrow, hardest-case work like long-horizon autonomous biology research1118.

Why It Matters

Opus 5 marked the moment the reasoning-model market shifted from selling intelligence to selling intelligence per dollar. When a model at half the price triples the field on a saturation-resistant reasoning benchmark910, and matches its flagship sibling on expert-level multidisciplinary reasoning within a point17, the practical question for buyers stops being "which model is smartest" and becomes "which effort level, on which tier, solves my task for the least spend." For enterprise workloads the 50% saving compounds past $1,000 a month at heavy tiers11, and with Opus 5.5 now undercutting Opus 5's own price while beating Fable 5.1 across most benchmarks320, the deflationary curve in frontier reasoning shows no sign of flattening.