What are the strengths of this model?

Same $5/$25 price as Opus 4.8 with a large capability jump Native self-verification; Frontier-Bench v0.1 more than doubles Opus 4.8 Leads on knowledge work (GDPval-AA v2 1,861 Elo) and science

What are the weaknesses of this model?

Anthropic's system card says it is not more capable overall than Fable 5 Loses DeepSWE v1.1 to GPT-5.6 Sol; trails Mythos 5 on cyber and legal work Verbose and slow; breaking change - extended thinking cannot be disabled at xhigh or max

What are the best use cases?

Long-horizon agentic coding and large-scale refactoring Root-cause analysis of hard-to-reproduce bugs Professional knowledge work and deep research

AnthropicProprietary

Claude Opus 5

Name: Claude Opus 5
Author: Anthropic

Compare this model

Anthropic's Opus-tier flagship, released July 24, 2026. It topped the Artificial Analysis Intelligence Index at 61 on launch day while keeping Opus 4.8's $5/$25 pricing, and pairs a 1M-token context window with 96.0% on SWE-bench Verified and 90.4% on ARC-AGI-2.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

License

Proprietary

Release Date

2026-07-24

Benchmark Performance

AA Intelligence Index

—

LMArena Elo

—

HLE

64.7

ARC-AGI-2

90.4

SWE-bench Verified

96.0

GPQA Diamond

93.2

MMLU-Pro

—

LiveCodeBench

—

AIME 2025

—

MATH-500

—

API Pricing

Input Price (per 1M tokens)

Output Price (per 1M tokens)

$25

Billing Mode: standard

Strengths

•Same $5/$25 price as Opus 4.8 with a large capability jump
•Native self-verification; Frontier-Bench v0.1 more than doubles Opus 4.8
•Leads on knowledge work (GDPval-AA v2 1,861 Elo) and science

Weaknesses

•Anthropic's system card says it is not more capable overall than Fable 5
•Loses DeepSWE v1.1 to GPT-5.6 Sol; trails Mythos 5 on cyber and legal work
•Verbose and slow; breaking change - extended thinking cannot be disabled at xhigh or max

Use Cases

•Long-horizon agentic coding and large-scale refactoring
•Root-cause analysis of hard-to-reproduce bugs
•Professional knowledge work and deep research

Deep Analysis

Artificial Analysis Intelligence Index

#1 overall at max effort; Claude Fable 5: 60, GPT-5.6 Sol: 59

SWE-bench Verified

96.0%

Anthropic-reported; vals.ai independently measured 97.0%

SWE-bench Pro

79.2%

SWE-bench Multilingual 89.5%; DeepSWE v1.1 68.8% (loses to GPT-5.6 Sol)

ARC-AGI-2

90.4%

ARC-AGI-1 97.5%; ARC-AGI-3 30.2%, roughly 4x the next-best model

Input / Output Price

$5 / $25 per 1M

Identical to Opus 4.8; half the cost of Fable 5 ($10/$50)

Context Window

1M tokens

128K max output; 300K via Batch API beta; knowledge cutoff May 2026

Strengths

・Tops the Artificial Analysis Intelligence Index at 61 while charging exactly what Opus 4.8 charged ($5/$25 per 1M) - a capability jump with no price increase, and roughly half of Claude Fable 5.
・Native self-verification. Opus 5 checks its own work without being told to, and Frontier-Bench v0.1 (43.3%) more than doubles Opus 4.8's 18.7% on long-horizon agentic coding.
・Strong on knowledge work and science: 1,861 Elo on GDPval-AA v2 (over 100 points clear of Fable 5 and GPT-5.6 Sol), +10.2 points over Opus 4.8 on organic chemistry and +7.7 on protein tasks.

Weaknesses

・Anthropic's own system card states Opus 5 "is not more capable overall than Claude Fable 5" - the index win is a composite result, not a clean sweep across every task.
・It loses DeepSWE v1.1 to GPT-5.6 Sol by roughly four points, and Anthropic notes it still trails Mythos 5 on cybersecurity and on legal work.
・Breaking API change: extended thinking is on by default and cannot be disabled at xhigh or max effort (a disabled-thinking request returns HTTP 400). Prompt scaffolds tuned for Opus 4.7/4.8 - especially "verify your work" instructions - can now degrade output by pushing the model toward over-checking.

Competitor Comparison

Model	Arena	SWE	GPQA	Price
Claude Fable 5	60 (AA Intelligence)	80.3% (SWE-bench Pro)	92.6%	$10/$50 per 1M
GPT-5.6 Sol	59 (AA Intelligence)	64.6% (SWE-bench Pro)	94.6%	$5/$30 per 1M
Claude Opus 4.8	N/A	18.7% (Frontier-Bench v0.1)	N/A	$5/$25 per 1M

Overview

Claude Opus 5 is Anthropic's flagship Opus-tier model, released on July 24, 2026 as the fourth Claude 5 model in under two months (after Mythos 5, Fable 5 and Sonnet 5). The pitch is not raw capability but price-to-capability: Anthropic says Opus 5 "comes close to the frontier intelligence of Claude Fable 5 at half the price," and the API rate confirms it - $5 per million input tokens and $25 per million output, exactly what Opus 4.8 cost, against Fable 5's $10/$50. The model ID is claude-opus-5, the context window is 1M tokens with 128K max output, and it shipped same-day across the Claude API, Claude.ai, Claude Code and Claude Cowork. It is the default on Claude Max and the strongest option on Claude Pro. Independent testing backed the launch claims immediately. On July 24, Artificial Analysis placed Opus 5 at #1 on both its Intelligence Index (61, the highest score it had ever recorded) and its Agentic Index (55.3), with a joint-first 78 on the Coding Index. Anthropic's own published figures are equally aggressive: 96.0% SWE-bench Verified, 79.2% SWE-bench Pro, 89.5% SWE-bench Multilingual, 90.4% ARC-AGI-2, 97.5% ARC-AGI-1, 93.2% GPQA Diamond, 70.6% OSWorld 2.0, 90.8% BrowseComp and a perfect 42/42 on IMO 2026. The two results worth singling out are Frontier-Bench v0.1 (43.3% against Opus 4.8's 18.7%) and ARC-AGI-3 (30.2% against GPT-5.6 Sol's 7.8%), both of which measure the ability to keep working after the obvious approach fails. The honest caveat is that Anthropic does not actually call Opus 5 its most capable model. The system card says plainly that Opus 5 "is not more capable overall than our most capable general-access model, Claude Fable 5," and Anthropic still recommends Fable 5 for work a model might run autonomously for days. Both statements can be true at once: Fable 5 remains the nominal ceiling, while Opus 5 tops the leading independent composite index and wins several agentic and coding evaluations at half the token cost. For most teams the routing decision is straightforward - Opus 5 becomes the everyday default, and Fable 5 gets reserved for the genuinely hardest jobs.

Benchmarks & Performance

Detailed Comparison

### Head-to-head with the models it actually competes with | Feature | Claude Opus 5 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 | | :--- | :--- | :--- | :--- | :--- | | **Role** | Everyday flagship, value pick | Absolute ceiling, multi-day autonomy | OpenAI frontier tier | Previous Opus flagship | | **Price (in/out per 1M)** | $5 / $25 | $10 / $50 | $5 / $30 | $5 / $25 | | **Context window** | 1M tokens | 200K tokens | 1.05M tokens | 1M tokens | | **AA Intelligence Index** | 61 (#1) | 60 | 59 | Lower tier | | **Strengths** | Best intelligence-per-dollar; self-verification; ARC-AGI-3 and Frontier-Bench leadership; science (chemistry, protein) | Highest ceiling on the hardest, longest-running autonomous jobs; health tasks | Terminal work, DeepSWE, GPQA Diamond, cybersecurity | Proven, stable scaffolds; identical price | | **Weaknesses** | Verbose and slow; hallucinates slightly more than 4.8; argumentative with legacy prompt scaffolds | Twice the price; 200K context | Higher output price; trails on ARC-AGI-3 by ~4x | Roughly half of Opus 5's Frontier-Bench score at the same cost | **The routing decision.** Opus 5 costs the same as Opus 4.8, so there is no economic reason to stay on 4.8 - the only migration cost is prompt scaffolding, which does need attention. Against Fable 5 the trade is explicit: Opus 5 gets you close to frontier intelligence at half the token price, and Anthropic itself recommends escalating to Fable 5 only for work a model might run autonomously for days. Against GPT-5.6 Sol the two models are close enough on coding to be a tie, so the deciding factors are output price ($25 vs $30 per 1M), whether your workload leans on ARC-AGI-3-style novelty (Opus 5 by a wide margin) or on terminal and cybersecurity work (Sol). **One asterisk worth repeating:** Anthropic's system card explicitly declines to call Opus 5 more capable overall than Fable 5. If a task is genuinely at the edge of what any model can do, the composite index ranking is not the number to route on.

Community Feedback

Reception on launch day was unusually positive for a mid-cycle release, largely because the price stayed flat. The recurring line across developer forums was that a same-price capability jump removes the usual "is the upgrade worth it" calculation entirely. * **What people liked.** The self-verification behaviour drew the most comment. Anthropic explicitly advises deleting "verify your work" style instructions carried over from older models, because Opus 5 now does this natively and the extra instruction pushes it toward over-checking. Several reviewers called this the single most actionable item for anyone with an established prompt library. Independent confirmation from Artificial Analysis on day one - #1 on both the Intelligence and Agentic indices - was also widely cited as stronger evidence than a vendor benchmark table. * **What people pushed back on.** Three complaints recur. Opus 5 is **verbose and slow** at high effort levels. It **hallucinates slightly more than Opus 4.8** in some reviewers' testing. And it is **argumentative with existing scaffolds** - prompts tuned for 4.7 or 4.8 can actively degrade its output, which caught several teams off guard during migration. The extended-thinking breaking change (HTTP 400 when thinking is disabled at xhigh or max) generated a round of bug reports before it was understood as intended behaviour. * **The ARC-AGI-3 skepticism.** A near-4x lead over the next-best model is unusual enough that several commentators flagged it. The reasonable reading is that ARC-AGI-3 is a research benchmark measuring novel problem-solving with no memorised template - it correlates with debugging genuinely strange production bugs, not with writing another CRUD endpoint, where a much cheaper model is fine. * **Practical adoption.** The emerging consensus is to make Opus 5 the default on Claude Max and Pro (where Anthropic already made it the default), reserve Fable 5 for multi-day autonomous runs, and audit prompt libraries before switching rather than after.

Use Cases

### Where Claude Opus 5 earns its price **1. Long-horizon agentic coding** * **Example:** An agent asked to migrate a service across a framework major version, where the first three approaches fail and the model has to diagnose why. * **Why Opus 5:** Frontier-Bench v0.1 measures exactly this - continuing productively after the obvious path breaks - and Opus 5's 43.3% more than doubles Opus 4.8's 18.7% at the same token price. Combined with 96.0% SWE-bench Verified and 89.5% SWE-bench Multilingual, it is the strongest available option for repositories that are not English-only Python. **2. Hard debugging and root-cause analysis** * **Example:** A flaky distributed-systems bug with no obvious reproduction, where the failure mode has no template in any training set. * **Why Opus 5:** ARC-AGI-3 at 30.2% versus GPT-5.6 Sol's 7.8% is the clearest signal any current benchmark gives about performance on problems with no memorised pattern. This is the workload where paying Opus rates is defensible. **3. Professional knowledge work and deep research** * **Example:** Turning messy source material - interview notes, filings, spreadsheets - into a defensible analyst report. * **Why Opus 5:** GDPval-AA v2 at 1,861 Elo is more than 100 points clear of both Fable 5 and GPT-5.6 Sol, and AA-Briefcase at 1,720 Elo points the same direction. The 1M-token context window, which Anthropic claims now stays stable across its full length, matters here more than anywhere else. **4. Scientific research, especially chemistry and biology** * **Example:** Reasoning over reaction mechanisms or protein-function questions as part of a literature-review agent. * **Why Opus 5:** Anthropic positions it as its most capable generally available model for scientific research, reporting +10.2 points over Opus 4.8 on organic chemistry and +7.7 on protein tasks. GPQA Diamond at 93.2% supports the general claim, though GPT-5.6 Sol edges it there at 94.6%. **5. Computer use and desktop automation** * **Example:** An agent driving a browser and desktop applications to complete a multi-step operational workflow. * **Why Opus 5:** OSWorld 2.0 at 70.6% beats Fable 5's best result at just over a third of the cost per task, and Zapier's AutomationBench pass rate is roughly 1.5x the next-best model at the same cost. **When to pick something else** * **Claude Fable 5** for work a model may run autonomously for days, and for health tasks where it still leads. * **Claude Mythos 5** for legitimate security research and legal work, where Anthropic acknowledges Opus 5 trails. * **GPT-5.6 Sol** for terminal-heavy pipelines and DeepSWE-style workloads. * **Claude Sonnet 5** for high-volume production traffic where Opus-tier reasoning is not required and the price difference compounds.

Latest News

**Launch (July 24, 2026).** Anthropic released Claude Opus 5 as the successor to Opus 4.8 and the fourth Claude 5 model in under two months, following Mythos 5, Fable 5 and Sonnet 5. It went live the same day on the Claude API (model ID `claude-opus-5`), Claude.ai, Claude Code and Claude Cowork, and became the default model on Claude Max and the strongest option on Claude Pro. **Pricing held flat.** $5 per million input tokens and $25 per million output - identical to Opus 4.8 and half of Claude Fable 5's $10/$50. Prompt caching is priced at $6.25/1M write and $0.50/1M read on the 5-minute TTL, with a 1-hour TTL at $10.00/1M write. A **fast mode** runs at roughly 2.5x default speed for twice the base rate ($10/$50), available on the Claude Platform and through usage credits in Claude Code. **Independent ranking on day one.** Artificial Analysis placed Opus 5 first on both its Intelligence Index (61 - the highest score it had recorded for any model) and its Agentic Index (55.3), with a joint-first 78 on the Coding Index, ahead of Fable 5 and GPT-5.6 Sol. **Two behaviour changes that break existing code.** Extended thinking is now **on by default** and can only be disabled at effort `high` or below; sending `thinking: {"type": "disabled"}` at `xhigh` or `max` returns a 400 error. Separately, Anthropic recommends **removing "verify your work" instructions** from prompts written for older models, because Opus 5 self-verifies natively and the duplicate instruction pushes it toward over-checking. **Longer context, claimed stable.** Anthropic says the 1M-token window holds quality across its full length, a direct response to documented reports of Opus 4.8 degrading past roughly 200K tokens. As of this writing there is no published independent long-context test confirming it, so it remains a vendor claim.

Independent testing backed the launch claims immediately. On July 24, Artificial Analysis placed Opus 5 at #1 on both its Intelligence Index (61, the highest score it had ever recorded) and its Agentic Index (55.3), with a joint-first 78 on the Coding Index. Anthropic's own published figures are equally aggressive: 96.0% SWE-bench Verified, 79.2% SWE-bench Pro, 89.5% SWE-bench Multilingual, 90.4% ARC-AGI-2, 97.5% ARC-AGI-1, 93.2% GPQA Diamond, 70.6% OSWorld 2.0, 90.8% BrowseComp and a perfect 42/42 on IMO 2026. The two results worth singling out are Frontier-Bench v0.1 (43.3% against Opus 4.8's 18.7%) and ARC-AGI-3 (30.2% against GPT-5.6 Sol's 7.8%), both of which measure the ability to keep working after the obvious approach fails.

The honest caveat is that Anthropic does not actually call Opus 5 its most capable model. The system card says plainly that Opus 5 "is not more capable overall than our most capable general-access model, Claude Fable 5," and Anthropic still recommends Fable 5 for work a model might run autonomously for days. Both statements can be true at once: Fable 5 remains the nominal ceiling, while Opus 5 tops the leading independent composite index and wins several agentic and coding evaluations at half the token cost. For most teams the routing decision is straightforward - Opus 5 becomes the everyday default, and Fable 5 gets reserved for the genuinely hardest jobs.

Sources

Analysis generated: 2026-08-11