Key Takeaways
- Claude Opus 5, released July 24, 2026, tops Artificial Analysis’s Intelligence Index (61) and Agentic Index (55.3), edging out Claude Fable 5 (60 / 52.8) and GPT-5.6 Sol (59 / 54.0) while running at half Fable 5’s API price.
- It keeps Opus 4.8’s pricing — $5 per million input tokens, $25 per million output tokens — against Fable 5’s $10/$50, and ships with a native 1-million-token context window.
- Independent benchmarks show Opus 5 pulling ahead on agentic knowledge work and frontend coding (AA-Briefcase, GDPval-AA v2, WebDev Arena), but the gains come from a five-tier adaptive effort dial, not from using fewer tokens — Opus 5 is measurably more verbose and slower per task than its predecessor.
- Anthropic calls it the company’s most aligned Opus model to date, though it still trails the restricted Claude Mythos 5 on offensive cybersecurity and biological research.
Anthropic has released four models in roughly two months — Mythos 5, Fable 5, and Sonnet 5 in June, followed by Claude Opus 5 on July 24. That pace alone would be notable. What makes Opus 5 the one worth stopping for is the specific bet it represents: rather than chasing a new intelligence ceiling, Anthropic built a model that gets close to its own frontier system, Fable 5, at half the sticker price, and backed that claim with independently verified benchmark data within days of launch. For a market where enterprise buyers have grown openly critical of ballooning AI bills, that’s the story that matters more than another leaderboard trophy.
The Benchmark Matrix
The headline number comes from Artificial Analysis, the third-party evaluator that tracks intelligence-versus-cost tradeoffs across the industry. At maximum reasoning effort, Opus 5 scores 61 on the Artificial Analysis Intelligence Index v4.1, a composite built from nine evaluations spanning coding, science, and factual reasoning. That puts it one point above Fable 5 (60) and two points above GPT-5.6 Sol (59) — a narrow gap, but enough to make Opus 5 the nominal leader among more than 160 tracked models. It’s also a substantial jump from Opus 4.8’s own score of 56 on the same index just two months earlier.
The more revealing number sits on Artificial Analysis’s Agentic Index, which weights benchmarks toward autonomous, multi-step task performance rather than single-turn reasoning. There, Opus 5 leads more clearly at 55.3, ahead of GPT-5.6 Sol (54.0) and Fable 5 (52.8) — and well ahead of Opus 4.8’s prior score of 47.2. That eight-point jump in two months is the number Anthropic will lean on hardest in enterprise sales conversations, since agentic reliability, not static knowledge recall, is what most production deployments actually need.
Two further benchmarks reinforce the pattern. On GDPval-AA v2, Artificial Analysis’s evaluation of realistic professional knowledge work, Opus 5 at max effort scores 1,861 Elo — more than 100 points ahead of both Fable 5 and GPT-5.6 Sol. On AA-Briefcase, the firm’s proprietary test for agentic knowledge tasks involving large private file sets, Opus 5 scores 1,720 Elo, a 146-point lead over Fable 5, while costing $17.79 per completed task versus Fable 5’s $22.30 — a 20% reduction. Artificial Analysis found that even Opus 5’s mid-tier “high” effort setting beats Fable 5’s AA-Briefcase score while costing less than half as much per task.
Not every result favors Opus 5. On AA-Omniscience, a factual-knowledge benchmark, Opus 5 improves seven points over Opus 4.8 but still lags Fable 5 — a gap Artificial Analysis attributes to the models’ different size classes. More notably, Opus 5’s hallucination rate on that same benchmark climbs 14 points to 50%, because the model answers more often instead of declining when it’s uncertain. That’s a real tradeoff, not a rounding error, and it’s worth factoring into any deployment where factual precision outweighs task completion.
Coding and Agentic Performance
Software engineering is where Opus 5’s positioning gets most concrete. On SWE-bench Verified, it scores 96.0%, currently the top result among tracked models. On Terminal-Bench v2.1, which evaluates agents operating autonomously in real terminal environments, it reaches 89% at max effort — roughly matching the previous leader, GPT-5.6 Sol. On the Artificial Analysis Coding Agent Index, Opus 5 paired with Claude Code shares first place with GPT-5.6 Sol paired with Codex, and separately posts the top score on the SWE-Atlas-QnA sub-benchmark.
Frontend work tells a similar story through a different lens. On Arena.ai’s WebDev leaderboard — which pits models head-to-head on live, rendered web-app builds judged by human voters — claude-opus-5-max currently sits at the top with an Elo of 1,725, ahead of Moonshot’s kimi-k3-max at 1,682, across a leaderboard that has logged nearly half a million votes. On Arena.ai’s separate Agent Arena, which scores models on millions of real-world, long-horizon agentic sessions using tool orchestration, Opus 5 Max lands at #2, trailing only Anthropic’s own Fable 5, with Opus 5 High close behind at #3.
That “trailing only Fable 5” pattern shows up repeatedly. Anthropic is effectively competing against itself at the top of several leaderboards, which is arguably the more interesting market dynamic than any single benchmark score: the company now has multiple models occupying the top tier simultaneously, differentiated by cost and latency rather than by a wide capability gap.
The Economics: Pricing and the Effort Dial
Opus 5 is priced at $5 per million input tokens and $25 per million output tokens — identical to Opus 4.8, and exactly half of Fable 5’s $10/$50 rate. That pricing decision is the core of Anthropic’s pitch: buyers get a model that closes most of the gap to Fable 5’s frontier intelligence without paying the frontier premium.
But the efficiency story is more nuanced than “same price, less work.” Artificial Analysis’s task-level data shows Opus 5 is not a leaner model than its predecessor — it’s a more thorough one. At max effort, it averages 36.2 minutes per AA-Briefcase task, around 50% longer than Opus 4.8’s 24.1 minutes, driven mainly by a jump in turn count: 103 turns at max effort versus 55 for Opus 4.8. Separately, Artificial Analysis’s own model page describes Opus 5 as “somewhat verbose” to “very verbose” depending on effort tier, generating more output tokens per Intelligence Index run than the median model tested. Cost per Intelligence Index task lands at $2.03 for Opus 5 at max effort — cheaper than Fable 5 with fallback ($2.75), but still more expensive than Opus 4.8 ($1.80) or Sonnet 5 ($1.53) on the same measure.
In other words, Opus 5’s cost advantage comes from where it sits relative to Fable 5, not from spending fewer tokens than the model it replaces. The real lever for controlling spend is the effort setting itself, which is where Opus 5’s architecture actually earns its “extended thinking” framing.
Extended Thinking Architecture: How the Effort Dial Works
Opus 5 ships with adaptive thinking enabled by default, using a five-tier effort scale — low, medium, high (the default), xhigh, and max — that determines how much internal deliberation the model applies before answering. Unlike the older fixed budget_tokens approach, adaptive thinking lets Claude evaluate each request’s complexity in real time and allocate reasoning tokens accordingly, rather than following a static budget set in advance.
One architectural detail is worth flagging for anyone migrating from earlier models: on Opus 5, extended thinking cannot be fully disabled at the two highest effort tiers. Thinking can be turned off at effort levels of “high” and below, but attempting to disable it at “xhigh” or “max” returns an error — a change from Opus 4.8, where omitting a thinking configuration simply ran the request without reasoning at all.
This matters for a second reason: Anthropic’s documentation notes that a request’s max_tokens setting on Opus 5 covers thinking tokens and visible response text together, whereas the two were more separable on Opus 4.8. Teams migrating existing token limits from the previous model can find that the same numeric ceiling now constrains a different — and often smaller — share of usable output, since a larger portion is consumed by reasoning before the model produces its final answer.
The self-correction behavior the model is known for shows up most clearly in edge-case testing rather than everyday use. In Anthropic’s own evaluation against the 2026 International Mathematical Olympiad problem set, Opus 5 was run with adaptive thinking set to max and asked to produce rigorous, self-contained proofs with no external tools, inside a 256,000-token output ceiling. When an attempt exhausted that limit without finishing, the system automatically resampled the problem at a lower effort level rather than simply failing — a practical illustration of the “dynamic self-correction” Anthropic has built into the model’s operating loop, rather than a claim about reduced path divergence on ordinary tasks. Independent reporting has separately described Opus 5 as a model that “verifies its own work and recovers from errors without intervention” and “requires less back and forth” than earlier Claude models, though that framing comes from Anthropic’s own characterization at launch rather than a benchmarked metric.
Context Window and Long-Horizon Work
Opus 5 carries a native 1-million-token context window with a maximum output of 128,000 tokens — the same context ceiling as Fable 5, and large enough to hold multi-file codebases, long document sets, or extended agentic session histories without external retrieval. Combined with the longer turn counts described above, that context budget is what makes multi-file refactoring and extended coding sessions viable in a single continuous session rather than a chain of separately-primed calls. Two smaller platform changes shipped alongside Opus 5 reinforce that use case directly: developers can now change which tools are available to Claude mid-conversation without invalidating the prompt cache, and API customers can enable automatic fallbacks, so a request flagged by a safety classifier reroutes to another available model instead of being blocked outright.
Safety, Alignment, and the Mythos Ceiling
Anthropic describes Opus 5 as its most aligned Opus model to date and “the least susceptible to being tricked into misuse,” according to a company statement. An internal behavioral audit reportedly found higher adherence to Claude’s Constitution and fewer incidents of deceptive behavior compared with previous models.
The model’s safety posture is also deliberately capped relative to Anthropic’s restricted Mythos-class systems. Opus 5 trails Claude Mythos 5 on cybersecurity and biological-research evaluations, and its safety classifiers are calibrated accordingly: it can identify vulnerabilities in source code, but its guardrails block binary-based vulnerability scanning, penetration testing, and exploit generation outright. Anthropic says those classifiers intervene roughly 85% less often than the equivalent controls on Fable 5, reflecting a narrower but still present set of restrictions rather than an unrestricted cybersecurity capability. Existing members of Anthropic’s Cyber Verification Program get access to a version of Opus 5 with fewer restrictions for approved security work; in Claude.ai, Claude Code, and Claude Cowork, requests flagged by those classifiers fall back to Opus 4.8 by default.
What Developers and Enterprises Are Saying
Early technical discourse has centered less on raw intelligence and more on the price-to-performance shift. With Opus 5 holding Opus 4.8’s $5/$25 rate while closing most of the gap to Fable 5’s $10/$50 frontier tier, developers building continuous or high-frequency agentic loops — coding agents left running across a workday, background research pipelines, IDE-integrated refactoring tools — get a meaningfully cheaper option than Fable 5 without dropping to Sonnet-tier capability. That framing lines up with Opus 5’s strong showing on Arena.ai’s WebDev and Agent leaderboards, both of which reward exactly that kind of sustained, tool-using performance rather than single-shot benchmark answers.
At the same time, practitioner guidance published shortly after launch cautions against defaulting straight to maximum effort. Anthropic’s own benchmark data shows coding performance peaking at medium effort on several tasks and degrading past that point, meaning the “always use more thinking” instinct from earlier reasoning models doesn’t transfer cleanly to Opus 5’s five-tier system. The practical recommendation circulating among early adopters is to start at medium effort for day-to-day coding work and reserve high or xhigh for tasks that have been tested and shown to benefit from the extra deliberation.
FAQ
Is Claude Opus 5 more expensive than Claude Opus 4.8? No. Opus 5 keeps the exact same pricing as Opus 4.8: $5 per million input tokens and $25 per million output tokens. The intelligence and agentic benchmark gains come at no change in headline price.
How does Claude Opus 5 compare to Claude Fable 5? Opus 5 scores close to Fable 5 on most static benchmarks (61 vs. 60 on the Artificial Analysis Intelligence Index) and ahead of it on agentic and coding evaluations, while costing half as much per token ($5/$25 vs. $10/$50). Fable 5 still leads on factual knowledge (AA-Omniscience) and remains Anthropic’s recommended option for the longest, most autonomous multi-day tasks.
What is the “effort” setting on Claude Opus 5? It’s a five-level control — low, medium, high, xhigh, and max — that determines how much internal reasoning the model applies before responding. Higher effort generally improves results on complex tasks but increases both latency and token cost, and Anthropic’s own data shows some coding tasks peak at medium effort rather than max.
Does Claude Opus 5 use fewer tokens than previous models? Not consistently. Independent benchmark data shows Opus 5 taking more turns and running longer per task than Opus 4.8, and generating more output tokens on intelligence evaluations than the median tracked model. Its cost advantage comes from being priced well below Fable 5, not from using fewer tokens than its own predecessor.
Can Claude Opus 5 be used for cybersecurity research? With restrictions. It can identify vulnerabilities in source code, but built-in classifiers block binary-based scanning, penetration testing, and exploit generation. Members of Anthropic’s Cyber Verification Program can access a less-restricted version for approved security work.
Closing Analysis
The unresolved question isn’t whether Opus 5 is capable — the benchmark spread across Artificial Analysis, Arena.ai, and SWE-bench is consistent enough to take at face value. It’s whether “close to Fable 5 at half the price” holds up once real-world task diversity, not curated benchmarks, is factored in, particularly given Opus 5’s higher hallucination rate on uncertain queries and its longer, more turn-heavy execution style. Watch for how Anthropic’s effort-tier guidance evolves as more enterprise usage data comes in, and whether the model’s positioning as an “everyday” flagship holds once Mythos-class capability inevitably becomes the new frontier comparison point.
