Anthropic Ships Opus 5: Half The Price of Fable 5 — And Even Better

Anthropic released Claude Opus 5 on 24 July. The model narrowly tops the Artificial Analysis Intelligence Index, costs half as much as the company’s own flagship Fable 5 – and, according to the independent tests, hallucinates in half the cases where it is uncertain.

It is the fourth model Anthropic has shipped within two months. Claude Opus 5 has been available across all of the company’s platforms since launch day, and for developers as claude-opus-5 via the Claude API. In the subscription products it is the new default model on Claude Max and the strongest model available on Claude Pro. Its position within Anthropic’s own line-up is unusual: Opus 5 is not the most expensive model in the house, but in most measurements it is the most capable one.

Price: unchanged in absolute terms, halved in relative terms

API pricing stays exactly at the level of its predecessor Opus 4.8: 5 US dollars per million input tokens and 25 US dollars per million output tokens. For comparison: Claude Fable 5, launched in June, costs 10 and 50 dollars respectively. That makes Opus 5 half the price of the flagship.

Further terms, per Anthropic and Artificial Analysis:

• Caching: cache writes carry a 25 percent premium (6.25 dollars per million tokens, five-minute TTL), cache hits are discounted by 90 percent (0.50 dollars).
• Fast mode: roughly 2.5 times the default speed at twice the base price, on the Claude Platform and via usage credits in Claude Code.
• Context window: 1 million tokens, unchanged from Opus 4.8.
• Five effort settings (low, medium, high, xhigh, max), which let users trade intelligence against token consumption.
• No data retention requirement for general access – unlike Fable 5, where prompts are held for 30 days.

Economically, the effort dial is the real lever. According to Artificial Analysis, output token consumption varies by a factor of roughly eight between low and max, spanning 407 Elo points on the GDPval-AA v2 benchmark.

The benchmarks: first place, but barely

Artificial Analysis, an independent benchmarking provider, says it evaluated Opus 5 on Anthropic’s behalf ahead of the release – a fact the company discloses and one worth keeping in mind when reading the numbers. The Intelligence Index in its current version 4.1 aggregates nine individual tests: GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity’s Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR.

The gap at the top is smaller than the margin of error one would reasonably grant such composite indices. Within the Opus 5 family itself, the effort settings land at 61 (max), 60 (xhigh), 59 (high) and 56 (medium) – so the highest setting buys two points over “high” at substantially more tokens.

The leads are clearer wherever agentic work is measured. On GDPval-AA v2, Opus 5 (max) reaches 1861 Elo, 114 points ahead of Fable 5 and more than 100 ahead of GPT-5.6 Sol. On AA-Briefcase, the benchmark for agentic knowledge work with deliverables such as reports, presentations and spreadsheets, it scores 1720 Elo, +146 over Fable 5. On the Coding Agent Index, Opus 5 (xhigh) combined with Claude Code shares first place with GPT-5.6 Sol plus Codex (67 points each); on Terminal-Bench v2.1 Opus 5 hits 89 percent, roughly level with the previous leader.

On cost per task in the Intelligence Index, Opus 5 (max) averages 2.03 dollars – 26 percent below Fable 5 (2.75 dollars), but still above Opus 4.8 (1.80) and Sonnet 5 (1.53).

Anthropic’s own figures point the same way: on Frontier-Bench v0.1 more than a doubling over Opus 4.8 at a lower cost per task, on CursorBench 3.2 a result within 0.5 percent of Fable 5’s peak score at half the cost, and on ARC-AGI 3 a score the company puts at three times the next-best model. For its life sciences evaluations Anthropic reports a 10.2 percentage point lead over Opus 4.8 on organic chemistry and 7.7 points on protein-related tasks. These numbers come from internal runs and have not been independently replicated.

The weak spots

Factual knowledge and hallucinations. On AA-Omniscience, Opus 5 trails Fable 5 – expected, given the models sit in different size classes. Accuracy improves by seven points over Opus 4.8, but the model answers more often when uncertain: its hallucination rate climbs 14 points to 50 percent. For applications where a wrong but confidently phrased sentence gets expensive, that is the single most relevant figure in the whole release.

Speed and verbosity. At around 53 output tokens per second, Opus 5 is clearly slower than the median of 77. At max effort it generated 100 million tokens during the index evaluation against a median of 63 million – markedly more verbose. At “high”, by contrast, it is comparatively economical at 52 million tokens.

Price per token. Measured against all models tested, Opus 5 remains expensive: the median sits at 1.75 dollars input and 10 dollars output. The cost advantage does not come from the token price but from efficiency per task – and that, per Artificial Analysis, holds mainly at the higher effort settings. At lower settings, Opus 5 sits behind the GPT-5.6 family on the intelligence-versus-cost curve.

Gaps in scientific reasoning. On Humanity’s Last Exam, Opus 5 reaches 53 percent, roughly Fable 5 level. On CritPt, a frontier physics evaluation developed by researchers at Argonne and UIUC, it sits behind GPT-5.6 Sol, GPT-5.5 Pro and GPT-5.6 Terra.

Deliberate capability limits. Anthropic stresses that Opus 5 is intentionally not state of the art on risky, dual-use capabilities. At finding software vulnerabilities (OSS-Fuzz) it comes close to Mythos 5, the invitation-only model, but it lags well behind at developing exploits. Its cyber classifiers block binary-based vulnerability scanning, penetration testing and exploit generation, though the company expects them to trigger around 85 percent less often than Fable 5’s. When they do trigger in Claude.ai, Claude Code or Claude Cowork, the request falls back to Opus 4.8 by default – a design that raises the question of why a request deemed risky for one model is acceptable for another. On the biology side the reverse applies: blocked Fable 5 requests will now route to Opus 5, which Anthropic says makes it the most capable generally available model for scientific research.

Context

Anthropic is explicitly not selling Opus 5 as a capability jump at the absolute frontier, but as a price-performance argument: close to Fable 5 intelligence at half the price, built for everyday use rather than record scores. The market is visibly shifting from “who has the best model?” to “what does a completed task cost?” – a move that Moonshot’s Kimi K3 and the GPT-5.6 family also foreground in their launch messaging.

Two caveats remain. First, the human preference rankings (LMArena and related boards) have not yet rated Opus 5 – Fable 5 currently leads there. Second, a large share of the published figures comes either from Anthropic itself or from a pre-release evaluation commissioned by Anthropic. Independent re-measurement across multiple harnesses is still outstanding.