by Anthropic · Claude 5 family · best for default frontier model for agentic work
Claude Opus 5 is Anthropic's current Opus-tier model, released July 24, 2026 as the direct successor to Opus 4.8 and the first Opus in the Claude 5 generation. Anthropic calls it a step change rather than an increment: on its own launch table it roughly matches or beats Claude Fable 5 on coding, computer use and knowledge work while costing half as much, at the unchanged Opus rate of $5/$25 per million tokens. Anthropic's model guide now tells customers to start with Opus 5 for most workloads and move up to Fable 5.1 only when Opus 5 at higher effort falls short. The one-sentence take for a buyer: this is the model to standardize on for agentic coding and long knowledge-work tasks, provided you re-baseline costs for thinking being on by default.
high or below; disabling it at xhigh or max returns a 400 error. Both are breaking changes for code written against Opus 4.8.low, medium, high, xhigh, max); Anthropic says Opus 5 converts extra effort into better results more reliably than any earlier Opus, and that low and medium effort now hold quality at a fraction of the tokens.| Benchmark | Score | Source |
|---|---|---|
| Humanity's Last Exam | 56.3% | anthropic.com 2026-07-24T00:00:00.000Z |
| LMArena Elo | 1493 | lmarena.ai 2026-09-13T00:00:00.000Z |
| LMArena Coding Elo | 1533 | lmarena.ai 2026-09-13T00:00:00.000Z |
| SWE-bench Verified | 96% | anthropic.com 2026-07-24T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“Anthropic's own default for most workloads: near-Fable capability at the Opus price, with no 30-day retention hurdle. Easy to standardize on.”
Opus 5 is the model Anthropic now tells customers to start with, and the reasoning holds up for a platform decision. It matches or beats Fable 5 on most of Anthropic's published evaluations at half the price, it ships on the Claude API, Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, and unlike the Fable line it carries no 30-day retention requirement, so zero-data-retention contracts still apply. The strategic caveat is cadence: Fable 5.1 arrived five weeks later and GPT-6 Astra two days after that, so Opus 5 is the value anchor of the lineup rather than its ceiling. Priority Tier is not offered on this model.
“Opus 5 turned Fable-class output into a mainstream price point, which squeezes GPT-5.6 Sol and forces premium tiers to justify their cost.”
In market terms Opus 5 collapsed the gap between Anthropic's two tiers: on Anthropic's own table it leads Fable 5 on terminal coding, computer use, knowledge work and business automation at half the rate card. Against GPT-5.6 Sol, its most direct price peer, it wins SWE-bench Pro (79.2 vs 64.6) and OSWorld 2.0 (70.6 vs 62.6) while trailing on DeepSWE and ARC-AGI-2. The moat remains agentic coding and computer use. The exposure is at the top: GPT-6 Astra leads the ARC-AGI-3 and WebDev arena boards by wide margins, and Google's current Pro, Gemini 3.1 Pro, is still a preview that sits just below Opus 5 on the text arena and far behind on MathArena.
“Same $5/$25 line item as Opus 4.8, but thinking is now on by default and billed as output, so re-baseline per-task cost before migrating.”
The rate card is unchanged from Opus 4.8: $5 input, $25 output, $0.50 cache reads, $2.50/$12.50 batch, and the full 1M window at standard rates. Two changes matter for the budget. Thinking is on by default and thinking tokens bill as output, so workloads that ran without thinking on Opus 4.8 will produce more output tokens. Against that, Anthropic and several customers report fewer tokens, turns and tool calls per completed task, and low and medium effort hold up well. Fast mode doubles the rate. GPT-5.6 Sol lists at $4/$20 on promotion and Grok 4.7 at $2/$6, so Opus 5 is not the cheapest capable option.
“The model ID swap is one line, but thinking-on-by-default breaks content[0].text parsers. Fix that, drop old verify prompts, and it flies.”
Migration is a model-ID change plus two breaking behaviors. Responses can now start with thinking blocks, so code that reads the first content block as text must select blocks by type, and disabling thinking at xhigh or max effort returns a 400 error. Anthropic also advises removing the "add a verification step" instructions written for older models, because Opus 5 verifies its own work and over-verifies when told to. In return: MCP Atlas 85.8%, SWE-bench Pro 79.2%, mid-conversation tool changes that keep the cache warm, a 512-token cache minimum and a full low-to-max effort ladder. With thinking disabled it occasionally writes tool calls into plain text.
“It checks its own work, pushes back with reasons and finishes long jobs, though answers run longer and arena voters rate chat much like 4.7.”
For daily heavy use, Opus 5 is the default model on Claude Max and the strongest one on Pro, and it behaves like a more careful colleague: it verifies before handing work back, holds a reasoned design objection rather than folding, and completes long multi-step jobs without leaving stubs. Anthropic rates its latency as moderate, between Sonnet 5 and Fable 5.1. Two trade-offs stand out. Default responses and written deliverables run longer than Opus 4.8's. And on the LMArena text leaderboard its high-effort variant sits at 1493, inside the band of Opus 4.6 and 4.7 rather than above it. The web-development arena is where voters prefer it clearly.
“Most headline wins are Anthropic-run evals on brand-new benchmarks; independent boards confirm the math and ARC gains, not a chat leap.”
Most of the launch table is vendor-reported, run by Anthropic on benchmarks only months old (FrontierBench v0.1, OSWorld 2.0, AutomationBench, GDPval-AA v2), and several figures shifted between Anthropic's July and September posts. Independent confirmation exists but is narrower: ARC Prize verified 90.4% on ARC-AGI-2 and 30.2% on ARC-AGI-3, and MathArena's aggregate jumped from 44.2 to 74.1 over Opus 4.8. The launch claim of three times the next-best ARC-AGI-3 score no longer holds, since GPT-6 Astra posted 62.7%. The system card itself reports slightly more factual hallucination than Opus 4.8, overconfident answers, and out-of-scope edits at high effort. No independent speed data is published yet.
xhigh/max) and higher output-token counts for workloads that ran without thinking on Opus 4.8.The full research notes behind this review — verified against primary sources.
Anthropic does not disclose architecture type, parameter counts, expert counts, layer counts or attention design for any Claude model, so those fields are null here. What is disclosed: a 1M-token context window with no smaller variant, 128K synchronous output (300K through batch with a beta header), adaptive thinking that is on by default, and the current Claude tokenizer, introduced with Claude Opus 4.7 (Anthropic's docs put 1M tokens at roughly 555,000 words on it, against about 750,000 on earlier tokenizers). Thinking text is returned as a summary, and only when requested with display: "summarized"; the default is omitted. The system card describes Opus 5 as not more capable overall than Claude Fable 5, deployed under the same ASL-3 protections as Opus 4.8.
Coding (10.0): SWE-bench Verified 96.0% and SWE-bench Pro 79.2% (Fable 5: 80.0%), SWE-bench Multilingual 89.5%, SWE-bench Multimodal 59.4%, FrontierCode v1.1 53.4%, all vendor-reported. The LMArena web-development board (1687 at max effort, third overall) is the independent signal; the text-coding category is flat against Opus 4.8. Reasoning (9.8): HLE 56.3% without tools and 64.7% with tools, ARC-AGI-2 90.4% and ARC-AGI-3 30.2% verified by the ARC Prize Foundation. Math (9.7): Anthropic reports a 42/42 gold-level IMO 2026 result graded by a model panel and confirmed by human graders, and ArxivMath 90.8%; MathArena's independent aggregate puts Opus 5 (max) at 74.1, above Opus 4.8 (44.2) and GPT-5.6 Sol (67.4) but below Fable 5.1 (82.1) and GPT-6 Astra (88.0). Agentic and tool use (10.0): FrontierBench v0.1 43.3% against 33.8% for Fable 5, OSWorld 2.0 70.6%, MCP Atlas 85.8% (82.2% for Opus 4.8), AutomationBench 26.0%, BrowseComp 90.8%. Long context (9.5): 1M tokens at standard pricing; on ProgramBench, whose episodes run up to the full window, the hidden-test pass rate rises from 83% after one episode to 93% after five. No MRCR figure is published. Multilingual (9.0): evaluated on Global MMLU, MILU and INCLUDE with results shown only as charts; Anthropic's published per-language data for earlier Claude models shows 96-98% of English performance across major European and Asian languages. Vision (9.4) and document/OCR (9.3): Anthropic says Opus 5 closes much of the gap to Mythos 5 on chart reasoning (Chartography) and professional documents (GDP.pdf), strongest when it can crop and inspect images with tools; LMArena vision 1289, tenth overall. Creative writing (8.7): partners describe clearer, more concise analytical writing, but the text arena shows no preference gain over Opus 4.7. Instruction following (9.3): tops Anthropic's constitution-adherence audit, but over-verifies when given legacy verification prompts and makes out-of-scope edits at high effort on FrontierCode. Function calling (9.6): parallel tools, structured output, cache-preserving tool changes; with thinking disabled it can occasionally write a tool call into plain text. Safety calibration (9.0): Anthropic's most aligned model on its behavioral audit, with low over-refusal, offset by cyber classifiers and slightly higher factual hallucination than Opus 4.8. Real-time data (7.0): May 2026 cutoff; current information requires the web search tool.
| Benchmark | Score | vs Predecessor | vs Top Competitor | Source |
|---|---|---|---|---|
| SWE-bench Verified | 96.0% | +7.4 vs Opus 4.8 (88.6%) | Fable 5: 95.0% | Anthropic system card |
| SWE-bench Pro | 79.2% | +10.0 vs Opus 4.8 (69.2%) | Fable 5: 80.0%; GPT-5.6 Sol: 64.6% | Anthropic system card |
| SWE-bench Multimodal | 59.4% | +21.0 vs Opus 4.8 (38.4%) | Fable 5: 54.1% | Anthropic system card |
| FrontierBench v0.1 (Harbor run) | 43.3% | +22.2 vs Opus 4.8 (21.1%) | GPT-5.6 Sol: 34.4%; Fable 5: 33.8% | Anthropic system card |
| DeepSWE v1.1 | 68.8% | +9.8 vs Opus 4.8 (59.0%) | GPT-5.6 Sol leads: 72.7% | Anthropic system card |
| Terminal-Bench 4.0 | 52.3% | not reported for Opus 4.8 | Fable 5.1: 55.8%; GPT-5.6 Sol: 37.3% | Anthropic, Fable 5.1 launch |
| CursorBench 3.2.0 | 70.0% | not reported for Opus 4.8 | Fable 5.1: 73.4%; GPT-5.6 Sol: 67.2% | Anthropic, Fable 5.1 launch |
| Humanity's Last Exam (no tools) | 56.3% | +6.5 vs Opus 4.8 (49.8%) | Fable 5: 56.5%; Fable 5.1: 60.9% | Anthropic system card |
| Humanity's Last Exam (with tools) | 64.7% | +6.8 vs Opus 4.8 (57.9%) | Fable 5: 63.9% | Anthropic system card |
| BrowseComp | 90.8% | +6.5 vs Opus 4.8 (84.3%) | GPT-5.6 Sol: 90.4% | Anthropic system card |
| OSWorld 2.0 | 70.6% | +14.9 vs Opus 4.8 (55.7%) | GPT-5.6 Sol: 62.6% | Anthropic system card |
| MCP Atlas | 85.8% | +3.6 vs Opus 4.8 (82.2%) | not reported | Anthropic system card |
| GDPval-AA v2 (Elo) | 1861 | +268 vs Opus 4.8 (1593) | GPT-5.6 Sol: 1736 | Anthropic system card |
| AutomationBench | 26.0% | +9.0 vs Opus 4.8 (17.0%) | GPT-5.6 Sol: 18.1% | Anthropic system card |
| ARC-AGI-2 (max effort, verified) | 90.4% | +18.3 vs Opus 4.8 (72.1%) | GPT-6 Astra: 95.0%; GPT-5.6 Sol: 92.5% | ARC Prize |
| ARC-AGI-3 (high effort, verified) | 30.2% | +28.7 vs Opus 4.8 (1.5%) | GPT-6 Astra: 62.7%; GPT-5.6 Sol: 7.8% | ARC Prize |
| MathArena overall (max effort) | 74.1 | +29.9 vs Opus 4.8 (44.2) | GPT-6 Astra: 88.0; Fable 5.1: 82.1 | MathArena |
| LMArena Text Elo (high effort) | 1493 ± 4 | +12 vs Opus 4.8 high (1481) | Fable 5 high: 1506 | LMArena |
| LMArena Text Coding Elo (high effort) | 1533 ± 7 | level with Opus 4.8 high (1533) | GPT-6 Astra max: 1543 ± 23 | LMArena |
| LMArena WebDev Elo (max effort) | 1687 ± 7 | +128 vs Opus 4.8 high (1559) | GPT-6 Astra max: 1800 | LMArena |
Rows sourced to Anthropic are vendor-reported (max effort, averaged over five trials unless noted; competitor figures are Anthropic's compilation). Anthropic's September re-runs differ slightly from the launch numbers: HLE 56.6% without tools and 63.6% with tools, GDPval-AA v2 1824, AutomationBench 26.9%. OSWorld 2.0 scores from the August task release (75.4% partial, 39.6% strict) are not comparable with the launch figure. Anthropic also reports a 42/42 IMO 2026 score, graded by a panel of three models and confirmed by human experts. The Opus 4.8 SWE-bench Verified baseline is the catalog figure. FrontierBench v0.1 is the successor to Terminal-Bench 2.1 with a new task set, and Terminal-Bench 4.0 is a different version again, so neither is stored under the terminal_bench key, which holds Terminal-Bench 2.x scores for other models. GPQA Diamond, MMLU-Pro, AIME, MMMU, tau-bench, MRCR and the Artificial Analysis Index have no published Opus 5 figure and are null.
No independent throughput or time-to-first-token measurement for Opus 5 was available, so both speed fields are null. Anthropic rates its latency as "Moderate": slower than Sonnet 5, faster than Fable 5.1. Real latency depends mostly on the effort setting, because thinking is on by default. Customer reports quoted by Anthropic point in both directions: one trading-benchmark team saw under half Opus 4.8's latency with about a seventh of the reasoning tokens, while one Fable 5.1 early-access customer found Fable 5.1 about twice as fast as Opus 5. For latency-critical paths, fast mode gives up to 2.5x output speed at twice the price (research preview, Claude API only, waitlist). For latency-insensitive work, the Batch API halves the cost.
| Surface | Cost | Notes |
|---|---|---|
| API input | $5 / 1M tok | Unchanged from Opus 4.8; full 1M window at standard rates |
| API output | $25 / 1M tok | Thinking tokens bill as output |
| Cached input (read) | $0.50 / 1M tok | 0.1x base; 512-token minimum cacheable prompt |
| Cache write (5m / 1h) | $6.25 / $10 per 1M tok | 1.25x / 2x base |
| Batch (in/out) | $2.50 / $12.50 per 1M tok | 50% off both |
| Fast mode | $10 / $50 per 1M tok | Research preview, Claude API only |
| US-only inference | 1.1x all token prices | inference_geo: "us"; also US Data Zone on Microsoft Foundry |
| Web search tool | $10 / 1,000 searches | plus token costs |
| Image input | billed as input tokens | no separate per-image price |
| Direct UI | $20/mo Pro ($17/mo annual) · Max from $100/mo | Default model on Max; available on Pro; not on Free |
| Free tier | none | Claude Free plan does not include Opus |
| Rate limits | 1,000 RPM · 2M input TPM · 400K output TPM | Start tier; separate bucket from Opus 4.x; Priority Tier not supported |
Proprietary and API-only: no open weights, no self-hosting, no fine-tuning. It is available on the Claude API, Claude Platform on AWS, Amazon Bedrock (anthropic.claude-opus-5), Google Cloud Vertex AI and Microsoft Foundry, all from launch day. Fast mode runs only on the first-party Claude API. First-party routing is global by default; US-only inference costs 1.1x, and partner clouds set their own regional pricing (regional endpoints on partner clouds carry a 10% premium over global). Opus 5 is also supported in Claude Managed Agents, including the Dreams research preview.
Opus 5 is governed by Anthropic's Responsible Scaling Policy and deployed with the same ASL-3 protections as Opus 4.8. The system card treats it as having CB-1 but not CB-2 capabilities and finds it below the automated AI R&D threshold. Anthropic's behavioral audit rates it the company's most aligned model to date (overall misaligned-behavior score 2.3), with the lowest deception rates and the least susceptibility to misuse. Cyber safeguards are narrower than on Fable 5: source-code vulnerability discovery is allowed, while binary scanning, penetration testing and exploit generation are blocked. The Cyber Verification Program unlocks a less restricted version for vetted teams. Unlike Fable 5 and Fable 5.1, Opus 5 has no 30-day data retention requirement, so zero-data-retention arrangements apply, and API inputs are not used for training by default. HIPAA-ready API access is available. Anthropic's trust center lists SOC 2 Type 2, ISO 27001, ISO 42001, CSA STAR and HIPAA coverage. The system card also flags that the model sometimes states answers confidently when it is unsure, and hallucinates factual claims slightly more than Opus 4.8.
Official SDKs cover Python, TypeScript, Java, Go, Ruby, C# and PHP, plus the ant CLI. Opus 5 is first-class in Claude Code, the Claude Agent SDK and Claude Managed Agents, and it is the default model on Claude Max and the strongest option on Claude Pro. Launch partners quoted by Anthropic include Cognition (Devin), Cursor, Zapier, Lovable, Box, JetBrains and AWS Kiro. Usage signal is strong: its high-effort variant has over 42,000 votes on the LMArena text board less than two months after launch, and it sits in the top three of the WebDev arena.
Yes for agentic coding, computer use and long knowledge-work tasks: the price is identical and Anthropic reports large gains (SWE-bench Pro 69.2% to 79.2%, OSWorld 2.0 55.7% to 70.6%). Handle the two breaking changes first: thinking is on by default, and disabling it at xhigh or max effort returns a 400 error. Opus 4.8 stays available, with retirement not sooner than 2027-05-28.
It can. Thinking tokens bill as output and Opus 5 thinks by default, so workloads that ran without thinking on Opus 4.8 produce more output tokens. Try low or medium effort, where Anthropic reports strong quality at a fraction of the tokens, and compare cost per completed task rather than per request.
Anthropic's guidance is to move up when your evals on Opus 5 at higher effort still fall short. Fable 5.1's $0.25 cache reads narrow the gap on cache-heavy agent loops, but it requires 30-day retention unless you are eligible for zero retention.
Yes. Unlike Fable 5 and Fable 5.1, Opus 5 has no data retention requirement for general access, API inputs are not used for training by default, and HIPAA-ready API access is available.
Cyber classifiers allow source-code vulnerability discovery but block binary vulnerability scanning, penetration testing and exploit generation. In Claude apps, flagged requests fall back to Opus 4.8; on the API you can turn on automatic fallbacks (beta). Anthropic expects about 85% fewer interventions than on Fable 5, and vetted security teams can apply to the Cyber Verification Program.
Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud and Microsoft Foundry, with US-only inference at 1.1x on first-party. Anthropic rates latency as moderate; fast mode (research preview, Claude API only) delivers up to 2.5x output speed at $10/$50.
The reliable knowledge cutoff is May 2026. For anything later, enable the web search tool at $10 per 1,000 searches plus token costs.
The closest price peer at $4/$20, a promotional rate through at least 2026-11-21 (list $5/$30). On Anthropic's launch table Opus 5 leads on SWE-bench Pro (79.2 vs 64.6), OSWorld 2.0 (70.6 vs 62.6), GDPval-AA v2 (1861 vs 1736) and ARC-AGI-3 (30.2 vs 7.8, ARC Prize verified); Sol leads on DeepSWE v1.1 (72.7 vs 68.8) and ARC-AGI-2 (92.5 vs 90.4). Choose Sol for price and OpenAI tooling, Opus 5 for agentic coding and computer use.
The tier above at $10/$50, with cache reads cut to $0.25. Anthropic's September table puts it ahead of Opus 5 on Terminal-Bench 4.0 (55.8 vs 52.3), HLE without tools (60.9 vs 56.6), AutomationBench (31.4 vs 26.9) and CursorBench 3.2.0 (73.4 vs 70.0). It requires 30-day retention except for eligible zero-retention customers, and Anthropic still recommends Opus 5 as the starting point for most workloads.
OpenAI's most capable model at $10/$50, and the leader on the independent boards: ARC-AGI-3 62.7% vs 30.2% and ARC-AGI-2 95.0% vs 90.4% (ARC Prize), LMArena WebDev 1800 vs 1687, MathArena 88.0 vs 74.1. Opus 5 costs half as much per token, and its text-arena score (1493) is level with Astra's early 1480 ± 12.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-09-21