Claude Opus 5: Frontier Intelligence at Half the Price

Claude Opus 5 launched on July 24, 2026, and the framing Anthropic chose is unusual for a flagship release. The company did not lead with a new capability ceiling. It led with economics: Opus 5, in Anthropic’s words, “comes close to the frontier intelligence of Claude Fable 5 at half the price,” and it is “designed to be used every day.” At $5 per million input tokens and $25 per million output, it costs exactly what Opus 4.8 cost, and exactly half what Fable 5 costs.

That positioning is the story. For most of 2026 the frontier tier has been something teams rationed — a model you route to for the hard 5% of tasks while a cheaper workhorse handles volume. Opus 5 is Anthropic arguing that the rationing layer can collapse, because the model that wins the benchmark is now also the model that is cheap enough to leave switched on. It is the new default on Claude Max and the strongest model available on Claude Pro.

This analysis covers what actually shipped, the full benchmark table against Fable 5, Opus 4.8, and GPT-5.6 Sol, the price arithmetic that makes the “half” claim load-bearing, the four evaluations where Opus 5 does not come first, the behavioural shift Anthropic documents around self-verification, the safeguards and alignment results, and a routing take for teams deciding what to change on Monday. For the wider Claude line, our when-to-use-which guide across Sonnet 5, Opus 4.8, and Fable 5 is the companion piece this release reshuffles.

Opus 5 is available today across all platforms, with the API model ID claude-opus-5. Pricing is $5 per million input tokens and $25 per million output — the same rates Opus 4.8 has carried since May 28, 2026. Fast mode returns on the same terms as its predecessor: roughly 2.5× the default speed at twice Opus 5’s base price, available on the Claude Platform and through usage credits in Claude Code.

Two plan-level details matter more than they might appear. Opus 5 is the new default model on Claude Max, and the strongest model on Claude Pro — which means a large population of paying users gets moved onto it without changing a setting. And, consistent with prior Opus models, it carries no data retention requirements for general access. That last point is a real differentiator: Mythos-class models including Fable 5 come with a mandatory 30-day retention policy that zero-data-retention agreements do not override, which has been a live blocker for regulated buyers since June.

The effort setting is central to reading every chart Anthropic published. Rather than a single score per model, the launch post presents performance as a curve against effort — the control that lets you trade intelligence against token spend. Opus 5’s claims are frequently of the form “better performance at a given cost,” which is a statement about the whole curve, not a single peak. If your team has not yet operationalised effort as a routing variable, this release makes that overdue; we covered the mechanics when the control landed alongside Opus 4.8 and dynamic workflows.

Anthropic published a head-to-head table covering Opus 5, Fable 5, Opus 4.8, and GPT-5.6 Sol across twelve evaluation families. Every figure below is vendor-reported. Reproduced in full, including the four rows where another model wins — a completeness most launch coverage skips.

Read down the Opus 4.8 column and the generational story is unambiguous. Frontier-Bench doubles. ARC-AGI-3 goes from 1.5% to 30.2%. OSWorld 2.0 adds nearly fifteen points. GDPval-AA v2 adds 268 points and clears Fable 5 by 114. Anthropic’s claim that Opus 5 “more than doubles Opus 4.8’s performance at a lower cost per task” on Frontier-Bench is borne out by its own table.

Read across the Fable 5 column and the story is subtler. Opus 5 wins most rows, but the margins on the coding benchmarks are inside noise — 53.4% against 53.5% on FrontierCode is a tie in everything but typography. Anthropic is candid about this on CursorBench 3.2, where it says Opus 5 at max effort performs within 0.5% of Fable 5’s peak score at half the cost per task. The pitch was never that Opus 5 beats Fable 5 outright. It is that the gap has narrowed to the point where paying double stops making sense for most work.

The arithmetic is simple enough to state in one line: Fable 5 costs $10 per million input tokens and $50 per million output. Opus 5 costs $5 and $25. For a workload that spends $4,000 a month on Fable 5 tokens, the equivalent Opus 5 spend is $2,000 — before any efficiency gain, purely on list rates.

But per-token price is the least interesting half of the story, and Anthropic knows it. The launch post repeatedly frames results as cost per task rather than cost per token, because a model that finishes in fewer turns with fewer tool calls costs less even at identical rates. Customer reports in the announcement put numbers on that: one financial-modelling team reported averaging nine percentage points higher accuracy with a third fewer turns and tool calls and 60% less time; a legal team reported similar performance while generating 26% fewer tokens on average than Opus 4.8 at max reasoning; a trading firm reported roughly a seventh of the reasoning tokens and under half the latency of Opus 4.8.

Those are self-reported figures from named early-access customers, not audited benchmarks, and they should be read as directional. But the direction is consistent, and it compounds with the rate cut. If you are moving from Fable 5, the halved rate and the token reduction multiply. This is exactly the calculation our per-task and per-user cost framework exists to make repeatable — sticker price alone will mislead you on every release this year.

A launch table with no losses is a marketing document. This one has four, and they are the rows worth knowing if you are routing production traffic.

DeepSWE v1.1 goes to GPT-5.6 Sol at 72.7% against Opus 5’s 68.8% — a four-point gap on an agentic coding benchmark, and the only row where an OpenAI model leads outright. Notably Fable 5 also edges Opus 5 here at 69.7%, making this the clearest case where the “close to Fable 5” framing costs you something real. FrontierCode v1.1 is a statistical tie at 53.4% against 53.5%. Humanity’s Last Exam without tools goes to Fable 5 by two-tenths of a point, though Opus 5 retakes it with tools enabled at 64.7% against 63.9% — which tells you the gap is in raw recall rather than in tool-augmented reasoning.

The held-out Legal Agent Benchmark is the widest loss: 11.7% against Fable 5’s 13.3%. Absolute scores are low across the board here — GPT-5.6 Sol manages 2.5% — so this is a benchmark where every model is failing most tasks, and a 1.6-point spread on a hard held-out set is thin ground for a routing decision either way. And on HealthBench Professional, the comparison column is Mythos 5 at 66.0% rather than Fable 5, with Opus 5 at 59.8% behind even GPT-5.6 Sol’s 60.5%.

The pattern across all four: specialist and domain-heavy evaluations. Opus 5 dominates general agentic capability — terminal coding, computer use, novel problem-solving, business workflows — and gives ground on narrow professional domains. If your workload is legal-agent or clinical, the frontier tier above still buys you something. If it is engineering or operations, it increasingly does not.

The most useful part of the announcement is not a benchmark. It is the section describing what Opus 5 does differently, which Anthropic summarises as being “much stronger at verifying its work and iterating carefully until it succeeds.” Three documented examples make the behaviour concrete.

On one Frontier-Bench task, the model was given a drawing of a machine part and asked to write code rebuilding it as a 3D FreeCAD model — but was deliberately given no way to view the drawing. Rather than failing or guessing, Opus 5 wrote its own computer vision pipeline to extract the geometry from raw pixels, then reconstructed the part. Anthropic reports it did so repeatedly, and that no competing model with the same setup solved it in five attempts. In a second case, given a real bug in a popular open-source package manager, it found the root cause and fixed an edge case the community’s own patch had missed, while a competing model fixed the surface symptom and declared the bug resolved. In a third, an engineer at a trading firm used it to build a market data feed for a new exchange in one session; finding no live feed to validate against, the model built its own test harness to confirm it was parsing the exchange’s data correctly.

The common thread is a model that constructs its own verification apparatus when none is provided. That is the difference between an assistant you supervise turn by turn and one you can hand a multi-hour task. It is also the capability that makes longer autonomous runs economically sensible, because the failure mode of long agentic sessions has always been confident wrongness compounding unchecked.

Two customer observations in the announcement sharpen the point further. One reviewer described handing off a pull request and finding the model verified the branches, checked the template, and reasoned about test implications rather than rushing to publish — where older models “jumped ahead and got caught on our checks.” Another described the model pushing back on a proposed design and not folding when pressed: it narrowed its objection to a single design question and proposed a compromise. For teams running agents with reduced oversight, that kind of calibrated disagreement is worth more than a benchmark point.

Anthropic also flags stronger visual outputs — sophisticated visualisations and interactive illustrations — and a broad improvement in scientific research, with every life-sciences evaluation improving over Opus 4.8. The standouts are organic chemistry, where inferring molecular structures from spectroscopy data scores 10.2 percentage points higher than Opus 4.8 on Anthropic’s internal benchmark, and protein sequence-variation prediction at 7.7 points higher.

On Anthropic’s automated behavioural audit, Opus 5 scores 2.3 on overall misaligned behaviour — the lowest of its recent models. The company states it adheres to Claude’s Constitution better than Opus 4.8, Sonnet 5, or Fable 5, exhibits the lowest rates of deceptive behaviour, is the least susceptible to being tricked into misuse, and is its safest model yet at avoiding reckless actions with hard-to-reverse side effects. That last property is the one that matters operationally: it is the difference between an agent you let touch production and one you do not.

On dangerous capabilities, Anthropic’s position is that Opus 5 does not advance the frontier. Evaluated alongside private-sector and government partners, it remains behind Mythos 5 in both biology research and offensive cybersecurity. The nuance is worth stating precisely, because it is where the interesting engineering sits: Anthropic deliberately avoided training Opus 5 on cyber tasks, yet it improved substantially anyway as a by-product of general capability, and now comes close to Mythos 5 at finding vulnerabilities. It remains substantially behind on exploiting them. On the OSS-Fuzz evaluation, the two models identify vulnerabilities with similar success while Opus 5’s exploit-development score sits far behind.

On biology the change is a loosening. Because Opus 5 carries a similar safeguard suite to Opus 4.8 rather than Fable 5’s stricter one, it is now Anthropic’s most capable generally available model for scientific research — and biology-related requests blocked on Fable 5 now route to Opus 5 rather than Opus 4.8. Anthropic is explicit that limitations remain on long-running autonomous research tasks, where it considers the substantial biology risk to sit, and that Mythos 5 remains stronger for that work. If you have been living with classifier friction on Fable 5, this release is a material quality-of-life change; we wrote up the production patterns for handling it in our guide to fallback models and production resilience.

Alongside the model, Anthropic shipped two beta platform features that will matter more to builders than to end users.

Mid-conversation tool changes. Developers can now change which tools Claude can use within a conversation without invalidating the prompt cache. That sounds like plumbing; it is not. Until now, an agent that needed a different toolset for a later phase of a task either carried every tool from the start — paying context and confusion costs throughout — or paid a full cache invalidation to swap. Phase-based agents with narrow, stage-specific toolsets now become cheap to build.

Automatic fallbacks on the API. Requests flagged by safety classifiers on Opus 5, or on Fable 5, can now be routed automatically to another model rather than blocked. With fallbacks on, API requests always route to the best available model by default instead of failing. For anyone who has built retry logic around classifier refusals — a recurring tax on Fable 5 since launch — this removes a category of production error handling.

Anthropic also published a dedicated prompting guide for Opus 5. Given how much of this release’s value is expressed through the effort setting rather than raw capability, that guide is worth reading before you port prompts across from Opus 4.8 wholesale.

The routing consequences differ sharply depending on which model you are on today.

The broader signal is one we have been tracking across every frontier release this year: the competitive axis has moved from headline intelligence to cost per completed task. Anthropic did not claim Opus 5 is smarter than Fable 5. It claimed Opus 5 gets close enough for half the money, and built its entire chart set around performance-per-dollar curves rather than peak scores. When the vendor with the strongest model on the market argues its cheaper model is the right default, the era of rationing the frontier tier is ending. If your team wants help benchmarking this on your real workloads and rebuilding routing around per-task economics, that evaluation is where our AI transformation engagements begin.