Pular para o conteúdo principal

Claude Opus 5: The Field Guide

Intermediário

On 24 July 2026 Anthropic released Claude Opus 5 (claude-opus-5) — the fourth model in two months, after Sonnet 5 (30 June) and Fable 5 / Mythos 5 (9 June). The headline is boring on purpose: same $5 in / $25 out per MTok pricing as Opus 4.8, same 1M context, same 128k output cap. The interesting part is what those unchanged numbers now buy you: 44.4% on Frontier-Bench v0.1 vs Opus 4.8's 18.7% (more than double), 96.0% on SWE-bench Verified, and 30.16% on ARC-AGI-3 — roughly three times GPT-5.6 Sol's 7.78% and twenty times Opus 4.8's 1.52%. Opus 5 is now the default model for Claude Max and the strongest available on Claude Pro.

This page is the practical field guide: what the numbers actually mean for your stack, the two API contract changes that will 400-error a naive migration, when Opus 5 legitimately replaces Fable 5 (and when it doesn't), and the new xhigh / max effort tier.

What you'll learn
  • Know what Opus 5 is: 1M context, 128k output, $5 in / $25 out per MTok — same shape as Opus 4.8, materially better numbers
  • Understand the two API constraints that break naive migrations: thinking-disabled restrictions at high effort, and thinking-on-by-default eating your max_tokens budget
  • Read the benchmarks the way agent-team leads read them — Frontier-Bench, ARC-AGI-3, SWE-bench Pro, OSWorld 2.0 — and know which are meaningful for your workload
  • Do the honest pricing math: when Opus 5 at max effort beats Fable 5 on cost-per-solved-task, and when it doesn't
  • Pick the right effort tier — the new xhigh and max tiers earn their keep on Frontier-Bench and ARC-AGI-3, waste tokens elsewhere
  • Plan the Opus 4.1 → Opus 5 deprecation before the 5 August 2026 retirement date

The one-paragraph version​

Opus 5 is a pricing-neutral capability bump, not a repricing. Same $5 / $25 per MTok as Opus 4.8, same 1M-token context, same 128k max output on the synchronous Messages API, same reliable knowledge cutoff shifted to May 2026. What changed is what those tokens do: on Frontier-Bench v0.1 (Anthropic's own hardest coding-agent eval) Opus 5 scores 44.4% at xhigh effort vs Opus 4.8's 18.7% — more than double at the same list price. On ARC-AGI-3, novel-problem solving jumps from 1.52% to 30.16%. Two contract changes will break naive migrations: adaptive thinking is on by default and shares your max_tokens budget with the response, and thinking: {"type": "disabled"} now returns 400 at the new xhigh or max effort tiers. Opus 5 is the new default on Claude Max, the strongest model available on Claude Pro, and available via the Claude API, Bedrock (anthropic.claude-opus-5), Google Cloud (claude-opus-5), Microsoft Foundry, and Claude Platform on AWS.

What actually changed vs Opus 4.8​

Opus 4.8 shipped in April 2026 as the first Opus with a 1M context window and adaptive thinking on by default. Opus 5 keeps every one of those defaults, so on the wire the request/response shape is nearly identical. Four things moved:

  • The numbers. Frontier-Bench v0.1 went from 18.7% to 44.4% (xhigh) / 43.3% (max). SWE-bench Multimodal from 38.4% to 59.4%. OSWorld 2.0 from 55.7% to 70.57%. Zapier AutomationBench from 17.0% to 26.0%. ARC-AGI-3 from 1.52% to 30.16% at high effort.
  • A new max effort tier slots above xhigh, and thinking: {"type": "disabled"} no longer works at xhigh or max (400 error). The intent is that at those tiers you're explicitly paying for reasoning, so silently disabling it is now a bug rather than a preference.
  • Self-verification is now default behavior, not a prompted habit. Anthropic's migration guide explicitly warns you to remove manual "check your work" instructions — otherwise you get double-verification and burned tokens. This is the most common surprise in the first day of running Opus 5 code that used to run on Opus 4.8.
  • Prompt caching no longer invalidates on mid-conversation tool changes (beta). If your agent loop swaps its tool list per step, the cache prefix now survives; on Opus 4.8 you paid a fresh write every time.

The two 400-errors that will break migration​

Guided walkthrough1 of 4
  1. Opus 4.8 let you disable thinking at any effort level. Opus 5 only allows thinking: {"type": "disabled"} at high effort or below. Pairing it with xhigh or max is a hard error. If you have a wrapper that always disables thinking (some teams do this to keep responses terse), either drop effort to high or delete the thinking field and let adaptive thinking run.

Minimal safe migration — Opus 4.8 to Opus 5

# Before (Opus 4.8) — worked
response = client.messages.create(
  model="claude-opus-4-8",
  max_tokens=4096,
  extra_body={"effort": "max"},
  thinking={"type": "disabled"},           # allowed on 4.8 at any effort
  system="After answering, verify your work step by step.",
  messages=[...],
)

# After (Opus 5) — three edits
response = client.messages.create(
  model="claude-opus-5",
  max_tokens=16384,                        # thinking now shares this budget
  extra_body={"effort": "max"},
  # thinking={"type": "disabled"}          # 400 at xhigh/max — remove it
  system="Answer the question.",           # drop the verify instruction — Opus 5 self-verifies
  messages=[...],
)

The benchmarks that actually matter​

Anthropic published a lot of numbers on launch day. Three are load-bearing for real workloads, one is a marketing trophy, and two are worth reading carefully because they show where Opus 5 is not the top choice.

  • Frontier-Bench v0.1 — 44.4% (xhigh) vs Opus 4.8's 18.7%. This is Anthropic's hardest agentic-coding eval — long-horizon tasks that require the model to run tools, iterate, and self-correct. The 2.4× jump at the same price is the single strongest reason to migrate.
  • SWE-bench Verified — 96.0%. Near-ceiling. If you were using Opus 4.8 or Fable 5 for SWE-bench-style bug fixes, Opus 5 is a straight upgrade. But SWE-bench Verified is now a saturated benchmark — small differences don't discriminate models the way they used to.
  • OSWorld 2.0 — 70.57% vs Opus 4.8's 55.7%. Computer-use benchmark. If you're running Claude for Chrome, Claude Cowork's browser mode, or anything else that clicks around a real UI, this is the number to watch.
  • ARC-AGI-3 — 30.16% (high effort) vs GPT-5.6 Sol's 7.78% and Opus 4.8's 1.52%. The marketing trophy. ARC-AGI-3 tests novel-problem solving that resists memorization, and Opus 5 is roughly 3× the next best model. Meaningful signal if your workload has real novelty; less meaningful if it's variants of things the internet has already solved.
  • SWE-bench Pro — 79.2% vs Fable 5's 80.0%. Fable 5 still edges Opus 5 by 0.8pp on the harder SWE-bench Pro. Not enough to justify Fable 5's 2× price for most teams, but if you're in the top decile of code complexity and every point matters, note this.
  • Cybersecurity exploitation — Opus 5 sits behind Mythos 5. Anthropic explicitly says Opus 5 is not the top cyber model. If you're doing defensive-cyber work through Project Glasswing, Mythos is still the pick.

When Opus 5 replaces Fable 5 (and when it doesn't)​

The interesting question this launch forces: when do you actually need Fable 5's $10 / $50 pricing? Anthropic's own claim is that Opus 5 hits within 0.5% of Fable 5's peak on CursorBench 3.2 at half the cost, and matches Fable 5's OSWorld 2.0 result at "just over one-third of the cost." That's the strong version of the argument. The honest version:

  • Pick Opus 5 for agentic coding, computer use, business automation, life-sciences reasoning, and anything where cost-per-solved-task matters more than the last 1-3pp of quality. This is >90% of production workloads.
  • Pick Fable 5 when the workload is (a) long-running agents where the marginal run genuinely benefits from the last percent (Anthropic frames Fable 5 as "next-generation intelligence for long-running agents"), (b) SWE-bench Pro-style hard code tasks where you've measured a real gap, or (c) any workload where you're already inside the accuracy-vs-cost curve where 2× cost buys real quality.
  • Pick Mythos 5 for invitation-only defensive-cyber via Project Glasswing.

Honest cost-per-solved-task math

# Fable 5 on a hard agentic task
#   ~500k input tokens, ~50k output tokens
#   500 * $10/M + 50 * $50/M = $5.00 + $2.50 = $7.50 per run
#   Fable 5 solve rate: ~46% (Frontier-Bench v0.1, xhigh)
#   Cost per solved task: $7.50 / 0.46 = ~$16.30

# Opus 5 on the same task
#   Same token budget
#   500 * $5/M + 50 * $25/M = $2.50 + $1.25 = $3.75 per run
#   Opus 5 solve rate: ~44.4% (Frontier-Bench v0.1, xhigh)
#   Cost per solved task: $3.75 / 0.444 = ~$8.45

# Opus 5 wins on cost-per-solved-task by ~48% at near-parity quality.
# Break-even: Fable 5 would need to be 2x cheaper per successful run,
# i.e. Opus 5 would have to drop below ~23% solve rate for Fable 5 to win.

The new max effort tier — when it earns its keep​

Opus 5 introduces max as an explicit tier above xhigh. Adaptive thinking now spans low → medium → high → xhigh → max, and high is the default on the Claude API and Claude Code. Two observations from the launch benchmarks:

  • On Frontier-Bench v0.1, xhigh actually scores 44.4% vs max at 43.3% — max is not strictly better. It burns more tokens per attempt but the extra reasoning doesn't always convert on this eval.
  • On tasks that reward long deliberation — ARC-AGI-3, novel research problems, hard mathematical reasoning (IMO 2026 gold-medal level, 42/42) — the higher tiers do earn their keep.

The practical rule: default to high, move to xhigh when your eval shows measurable gains, only use max for workloads where you've verified reasoning depth matters more than throughput. Don't set max and hope; measure.

Life sciences, safety, and cyber posture​

Three numbers worth pinning:

  • Life sciences. Opus 5 is +10.2 percentage points over Opus 4.8 on organic chemistry and +7.7pp on protein prediction. If you're building agents for structural biology, bioinformatics, or synthesis planning, this is a real jump — not a within-noise improvement.
  • Prompt injection resistance. On the Gray Swan benchmark, attacker success drops from Opus 4.8's 5.5% to 2.0%. In Claude Cowork browser environments, auto-mode safeguards bring browser prompt-injection success to 0% in Anthropic's measurements. Meaningful for anyone shipping Claude-for-Chrome-style agents.
  • Cyber classifiers. Opus 5's cyber safeguards intervene ~85% less often than they do for Fable 5 — because Opus 5 is deliberately not the top cyber model, the intervention rate is dialed back. If you were seeing spurious safety refusals on legitimate security-research prompts on Fable 5, Opus 5 will feel noticeably less trigger-happy.

Deprecation timeline you need to plan around​

  • Opus 4.1 (claude-opus-4-1-20250805) retires 5 August 2026 — eleven days from Opus 5's launch. If any production surface still pins Opus 4.1, migrate this week.
  • Opus 4.5 / 4.6 / 4.7 / 4.8 remain available as "Legacy models" in the platform overview, but expect standard 12-month deprecation cycles. Treat Opus 5 as the migration target for anything new.
  • Opus 5's model ID is a pinned snapshot, not an evergreen pointer. Starting with the 4.6 generation, the dateless-ID format still means "fixed snapshot." A future claude-opus-5-1 would be a new ID, not a silent upgrade of claude-opus-5.

Try it — one command​

Run Opus 5 with the recommended defaults

# Python — Claude SDK
from anthropic import Anthropic
client = Anthropic()

response = client.messages.create(
  model="claude-opus-5",
  max_tokens=8192,                    # room for adaptive thinking + response
  extra_body={"effort": "high"},      # default; move to xhigh only if evals justify
  system="You are a senior engineer. Answer directly and iterate until the task is complete.",
  messages=[{"role": "user", "content": "..."}],
)

# Claude Code
claude --model claude-opus-5

# For fast mode (2.5x speed, 2x cost)
# model="claude-opus-5-fast"  # separate model ID, $10/$50 per MTok

Cheat sheet​

Pressione Enter ou Espaço para virar o cartão. Use as setas esquerda e direita para navegar entre os cartões.Termo exibido.
1 / 8

Check your understanding​

Check yourself

0/5
  1. You migrate a wrapper that pinned `model="claude-opus-4-8"`, `effort: "max"`, and `thinking: {"type": "disabled"}` to Opus 5. What happens?
  2. Your Opus 4.8 prompt ends with 'Now verify your answer step by step.' You swap to Opus 5 and see roughly double the thinking-token cost on hard prompts. Why?
  3. For a production agentic-coding workload where solve rate matters and cost is a real constraint, which model does Opus 5's launch data most strongly support?
  4. You set `effort: "max"` on every request expecting strictly better quality than `xhigh`. Anthropic's launch data suggests what?
  5. Which of these is the deprecation date you must plan around after Opus 5's launch?

Where to go next​

Sources & further reading​