Anthropic shipped its fourth new model in under two months on July 24, 2026. Claude Opus 5 lands just weeks after the turbulent debut of Claude Fable 5 and Claude Mythos 5, and it arrives with a pitch that’s less about raw intelligence and more about economics: near-frontier performance, at half the price of Anthropic’s most capable public model, positioned as the thing you actually reach for every day.
That’s a notable shift in framing for a company that spent most of 2024–2025 racing on pure capability. Here’s what Opus 5 actually is, what Anthropic’s own data shows, how outside reviewers are reading it, and where the marketing deserves a skeptical eye.
Where Opus 5 sits in Anthropic’s lineup
Anthropic’s current model stack, from cheapest to most capable, looks like this:
| Model | Tier | Positioning |
|---|---|---|
| Claude Haiku 4.5 | Entry | Fast, cheap, high-volume tasks |
| Claude Sonnet 5 | Mid | Agentic, cost-efficient default for most workloads |
| Claude Opus 5 | Upper-mid | New: daily-use professional and coding work |
| Claude Fable 5 | Frontier (public) | Longest, most ambitious autonomous projects |
| Claude Mythos 5 | Frontier (restricted) | Limited-availability, highest raw capability |
Opus 5 replaces Opus 4.8 as the workhorse of that stack. According to Anthropic’s own announcement, the model comes close to Fable 5’s intelligence on many tasks but stops short of it — and it’s explicitly not being pitched as a replacement for Fable 5 on the longest-running, highest-stakes autonomous work. On Frontier-Bench and GDPval-AA, two of Anthropic’s coding and knowledge-work evaluations, the company says Opus 5 is now state of the art among its generally available models, though it remains behind Mythos 5 specifically on cybersecurity tasks.
Pricing: same rate, better output
The headline economics are simple. Opus 5 costs exactly what Opus 4.8 cost: $5 per million input tokens and $25 per million output tokens. Anthropic’s pitch is that you get meaningfully more capable output for that same price, rather than a cheaper price for the same capability.
Here’s how that lines up against the rest of the current lineup, per Anthropic’s official pricing documentation:
| Model | Input (per MTok) | Output (per MTok) | Notes |
|---|---|---|---|
| Claude Haiku 4.5 | $1 | $5 | — |
| Claude Sonnet 5 | $2 (intro, through Aug 31, 2026) → $3 | $10 → $15 | Introductory pricing ends Sept 1, 2026 |
| Claude Opus 5 | $5 | $25 | Same as Opus 4.8 |
| Claude Opus 4.8 | $5 | $25 | Predecessor |
| Claude Fable 5 | $10 | $50 | Frontier public model |
| Claude Mythos 5 | $10 | $50 | Limited/trusted access only |
A Fast mode is also available for Opus 5, running at roughly 2.5x the default speed for double the base price ($10/$50 per million tokens) — matching the arrangement Anthropic already offers on Opus 4.8.
One notable addition alongside the model itself: an adjustable “effort” setting, which lets developers dial reasoning up or down to trade intelligence for speed and token spend. Several early-access partners specifically praised this — Anthropic says the model beats or matches Opus 4.8’s maximum-reasoning output while burning fewer tokens at lower effort settings.
What Anthropic says the benchmarks show
Anthropic’s own announcement leans heavily on cost-adjusted performance rather than peak scores alone — a deliberate framing choice worth flagging up front, since it’s the company measuring itself against its own predecessor. With that caveat, the headline claims from the official Anthropic announcement are:
- Frontier-Bench v0.1 (software engineering): Opus 5 surpasses all other Anthropic models and more than doubles Opus 4.8’s performance at a lower cost per task.
- CursorBench 3.2: at maximum effort, Opus 5 lands within 0.5 percentage points of Fable 5’s peak score, at half the cost per task.
- ARC-AGI 3 (novel problem-solving): Opus 5 scores roughly three times higher than the next-best model.
- Zapier AutomationBench (end-to-end business tasks): pass rate around 1.5x the next-best model at equivalent cost; even the lowest effort setting reportedly beats every competing model.
- OSWorld 2.0 (computer use): outperforms every other model at any given cost point, beating Fable 5’s best result at roughly a third of the price.
- Life sciences evaluations: improvements across the board versus Opus 4.8, with the biggest jumps on organic chemistry tasks like inferring molecular structure from spectroscopy data (+10.2 percentage points) and protein sequence-function prediction (+7.7 points).
Anthropic also shared a handful of qualitative demonstrations — the one getting the most attention online involves Opus 5 being handed a photo of a mechanical part with no direct way to “view” the image, and responding by writing its own computer vision pipeline to extract the geometry before rebuilding the part as a 3D CAD model. Anthropic says no competing model solved the same task in five attempts under identical conditions. As with any single anecdote in a launch post, that’s worth treating as an illustration of a capability rather than a representative average.
Independent coverage: convergence and caveats
Outside reporting on the launch is largely consistent with Anthropic’s framing, though several outlets sharpened the caveats. Bloombergreported that Anthropic positioned the model as approaching the capabilities of Fable 5 in many categories at half the price, tying the launch directly to rising cost sensitivity among enterprise buyers and intensifying competition from Chinese AI labs.
VentureBeat’s coverage was the most explicit about the competitive subtext, framing the release as a signal that the AI industry is shifting focus from raw capability toward the economics of everyday use. The outlet also relayed specific efficiency claims from early enterprise partners — legal AI company Harvey reported comparable output quality to Opus 4.8’s maximum-reasoning mode while using notably fewer tokens, and financial research firm Fundamental Research Lab cited higher accuracy on hard modeling tasks with fewer turns and less time spent per task.
MacRumors’ summary confirmed the same benchmark categories Anthropic emphasized — agentic coding, novel problem solving, knowledge work, computer use, and multidisciplinary reasoning — and was direct about the cybersecurity gap: Opus 5 finds vulnerabilities at a level close to Mythos 5, but is “substantially behind” on turning those vulnerabilities into working exploits.
SiliconANGLE dug into the guardrail changes, noting Anthropic’s own estimate that safety classifiers will trigger roughly 85% less often for Opus 5 than they did for Fable 5 — a meaningful practical change for teams whose legitimate security work has previously tripped Anthropic’s filters. CNBC, meanwhile, pulled a useful quote from Anthropic’s head of product management for research on the strategic logic behind the launch: enterprises want value, and a cheaper option only matters if it still delivers comparable results.
Fortune’s framing is worth noting for context: this isthe fourth model Anthropic has released in less than two months, following Mythos 5, Fable 5, and Sonnet 5 in June. That release cadence is unusually aggressive even by 2026 standards, and it’s difficult to separate from the competitive pressure both Bloomberg and CNBC pointed to.
Safety and alignment claims
Anthropic’s announcement devotes real space to alignment data, which is worth taking seriously rather than skipping past as boilerplate. The company’s automated behavioral audit — its internal pre-deployment testing process — reportedly scored Opus 5 as its most aligned model to date, with the lowest rate of deceptive behavior of any recent Claude model and the lowest overall misaligned-behavior score (2.3, versus higher scores for Opus 4.8, Sonnet 5, and Fable 5, per Anthropic’s own chart).
On dual-use risk, Anthropic is explicit that Opus 5 does not advance the frontier: it remains behind Mythos 5 on both biology research and offensive cybersecurity, according to evaluations the company says were conducted with private-sector and government partners. The company frames this as intentional — Opus 5, like Opus 4.8 before it, was not deliberately trained on cyber tasks, though general capability gains have pulled it closer to Mythos 5 on vulnerability discovery even as it lags well behind on exploitation.
The practical safeguard changes: Opus 5’s cyber classifiers block binary-based vulnerability scanning, penetration testing, and exploit generation, while permitting source-code vulnerability discovery — with flagged requests falling back automatically to Opus 4.8 in Claude.ai, Claude Code, and Claude Cowork. Biology-related requests previously blocked on Fable 5 now route to Opus 5 rather than Opus 4.8, making it Anthropic’s most capable generally available model for legitimate scientific research, with the caveat that long-running autonomous biology work is still considered Mythos 5’s territory.
Early customer reception
Anthropic published quotes from roughly two dozen partners at launch — a standard part of any model release, and worth reading as curated marketing rather than independent testimony. That said, the specificity of some of the claims is notable: Zapier’s CEO said Opus 5 topped the company’s internal AutomationBench leaderboard without a token-spend increase over prior Claude models, and reported that previous models didn’t pass; Opus 5 hit 100% on a churn-prevention workflow test. Devin-maker Cognition and Cursor both described the model as reaching near-Fable-5 performance at Opus-tier cost on their respective coding benchmarks. Box’s CTO cited double-digit percentage improvements on data analysis and due-diligence workflows specifically.
Worth flagging: these are all launch-day partners with existing commercial relationships with Anthropic, most of whom had early access explicitly to produce these testimonials. That doesn’t make the numbers false, but it does mean they should be read as best-case, cherry-picked results rather than an independent benchmark.
The honest read on the marketing
A few things are worth separating from Anthropic’s framing before you take this at face value:
- “Half the price of Fable 5” is true but slightly misleading as a headline. Opus 5 isn’t cheaper than its own predecessor — it’s priced identically to Opus 4.8. The “half price” comparison is against Fable 5, a different and more expensive tier. That’s a legitimate comparison, but it’s marketing that flatters the number.
- The cost-adjusted benchmark framing is Anthropic’s own methodology. Charting “performance per dollar” rather than peak performance is defensible, but it’s also the framing most favorable to a company launching a mid-tier model. Independent, apples-to-apples benchmarking from third parties (LMArena, Artificial Analysis, etc.) will be the more reliable signal once it accumulates.
- The cybersecurity gap is real and Anthropic is upfront about it — which is worth crediting. Rather than burying the Mythos 5 comparison, the official announcement leads with it in the safety section.
- Release cadence is a story in itself. Four model releases in under two months — following a period that included a temporary export-control suspension of Fable 5 and Mythos 5 access in June — suggests Anthropic is iterating faster than typical annual or biannual cycles, likely in direct response to competitive pressure from OpenAI, Google, and a widening field of Chinese labs.
Claude Opus 5 vs. Opus 4.8: what actually changed
| Category | Opus 4.8 | Opus 5 |
|---|---|---|
| Price | $5 / $25 per MTok | Same |
| Coding (Frontier-Bench) | Baseline | More than double the score, per Anthropic |
| Computer use (OSWorld 2.0) | Weaker at any cost point | Anthropic’s best at any cost point |
| Effort/reasoning control | Limited | New adjustable effort setting (low/medium/high/max) |
| Cyber guardrail friction | Baseline | ~85% fewer classifier interventions (Anthropic estimate) |
| Alignment audit score | Higher misalignment score | Lowest score (2.3) of recent models |
| Biology research capability | Baseline | Most capable generally available model for this use case |
| Default placement | — | New default on Claude Max; strongest model on Claude Pro |
FAQ
Is Claude Opus 5 better than Claude Fable 5? Not across the board. Anthropic positions Fable 5 as the stronger model for the longest, most ambitious autonomous projects, while Opus 5 is built for everyday professional and coding work at a lower cost. On some coding benchmarks Opus 5 comes within fractions of a percentage point of Fable 5’s peak score; on cybersecurity and the most demanding long-horizon tasks, Fable 5 (and Mythos 5) remain ahead.
How much does Claude Opus 5 cost? $5 per million input tokens and $25 per million output tokens via the API — identical to Opus 4.8’s pricing. A Fast mode is available at $10/$50 per million tokens for roughly 2.5x the speed.
Is Opus 5 available to regular subscribers, or API only? Both. It’s available on the Claude API, and it’s now the default model on Claude Max and the strongest model available on Claude Pro.
Does Opus 5 have weaker safety guardrails than Fable 5? Its cybersecurity classifiers trigger less often — Anthropic’s own estimate is around 85% less frequently — but this reflects looser restrictions on legitimate vulnerability-discovery work, not a broader relaxation. Exploit generation and binary-based scanning remain blocked, and flagged requests fall back to Opus 4.8 by default.
What’s the “effort” setting? A new adjustable parameter that lets developers trade reasoning depth for speed and token cost, ranging from low to max effort. Anthropic and several early partners report that even lower effort settings on Opus 5 can match or beat Opus 4.8’s maximum-reasoning output.
