TL;DR: Anthropic released Claude Opus 5 on July 24, 2026 at $5/$25 per million tokens — the same sticker as Opus 4.8 and exactly half of Claude Fable 5’s $10/$50. It posts 96.0% on SWE-bench Verified and 79.2% on SWE-bench Pro, within 0.8 points of Fable 5 on the harder benchmark. For almost every coding workload that was running on Opus 4.8 or burning Fable 5 credits, this is the new default.

	Claude Opus 5	Claude Fable 5	Claude Sonnet 5
Best for	Daily agentic coding, long-horizon refactors	The hardest frontier tasks, deep research	Volume work, cost-sensitive loops
Input / Output per 1M	$5 / $25 ($10/$50 fast mode)	$10 / $50	$2 / $10 intro → $3 / $15 from Sep 1
SWE-bench Pro	79.2%	80.3%	63.2%
SWE-bench Verified	96.0% (system card)	95.0%	85.2%
The catch	Longer outputs; thinking-disable rules changed	2x the price for +0.8 pts	Clearly behind on hard multi-file work

Honest take: Switch. Opus 5 makes Fable 5 a specialty tool — pay double only for frontier research tasks and the hardest architecture work, and stop routing everyday agent sessions through Opus 4.8, which now loses by ten points on SWE-bench Pro at the identical price. The one group that shouldn’t move blindly: anyone who runs with thinking disabled — there’s a new 400-error rule waiting for you.

What Anthropic shipped on July 24

Claude Opus 5 is the first Opus of the Claude 5 family, and it’s positioned exactly where the name suggests: below Fable 5, above everything else. The API model ID is claude-opus-5 — a dateless pinned snapshot, not an evergreen pointer — with anthropic.claude-opus-5 on Amazon Bedrock and claude-opus-5 on Google Cloud Vertex AI and Microsoft Foundry, all live day one. Opus 4.8 stays available on every platform.

The specs, verified against Anthropic’s model docs on July 27:

• 1M-token context window — and unlike some earlier models, 1M is both the default and the only size; there is no smaller variant or long-context surcharge tier
• 128K max output tokens (up to 300K on the Batch API with the output-300k-2026-03-24 beta header)
• May 2026 knowledge cutoff — the freshest of any Claude model; Fable 5 and Sonnet 5 both sit at January 2026
• Adaptive thinking on by default, steered by the effort parameter rather than a manual thinking budget
• Pricing: $5/$25 per million tokens, cache hits at $0.50/M, batch at $2.50/$12.50, and a research-preview fast mode at $10/$50 that’s API-only (not on Bedrock, Vertex, or Foundry)

One quieter improvement worth money: the minimum cacheable prompt dropped from 1,024 tokens to 512, so short system prompts that silently skipped the cache on Opus 4.8 now cache normally.

If the model felt familiar before launch, there’s a reason — a research model labeled “Honeycomb EAP” appeared briefly in Cursor’s model picker around July 9. That was Opus 5 in early access testing.

On the subscription side, Opus 5 is the default model on Claude Max and the strongest model available on Claude Pro, and it landed in Claude Code the same day. Anthropic is also pointing all Opus 4.1 users here: that model is deprecated and retires August 5, 2026.

The benchmarks, sorted by who measured them

The launch headline is 96.0% on SWE-bench Verified — Anthropic’s own number, averaged over five trials per the system card. The independent vals.ai leaderboard currently shows Opus 5 at 97.0%, slightly above the vendor claim, which is not the direction these gaps usually run. Either way, SWE-bench Verified is saturated: Fable 5 reports 95.0%, and at these levels the benchmark separates almost nothing.

The numbers that actually rank models in 2026:

• SWE-bench Pro: 79.2% — versus Fable 5’s 80.3%, Opus 4.8’s 69.2%, and Sonnet 5’s 63.2%. Ten points over its price-identical predecessor; 0.8 points behind the model that costs twice as much.
• Frontier-Bench v0.1: 43.3% — Anthropic’s new state-of-the-art claim, ahead of Fable 5’s 33.7% and more than double Opus 4.8. Vendor-run, brand-new benchmark; treat the ordering as more meaningful than the absolute values.
• ARC-AGI-3: 30.16% at high effort — this one is independently verified by the ARC Prize Foundation, and it’s the eyebrow-raiser: roughly four times the next model’s score (GPT-5.6 Sol at 7.8%) on a benchmark built to resist memorization.
• OSWorld 2.0: 70.57% — versus 55.7% for Opus 4.8 on the newer computer-use benchmark. (Don’t confuse this with the older OSWorld-Verified board, where Opus 4.8 scored 83.4% — different benchmark, different scale.)
• CursorBench: 66.7% at default effort — a statistical tie with Fable 5’s 66.5%, per Cursor’s own launch announcement, at half the cost per task.

Two honest gaps: Opus 5 doesn’t appear on the official tbench.ai Terminal-Bench 2.1 leaderboard as of July 27, where Fable 5 still leads at 88.0% — no Opus 5 score to report there yet. And Anthropic itself notes Opus 5 remains behind Mythos 5 on cybersecurity tasks, which is academic for most readers since Mythos 5 is invitation-only.

Using it in Cursor

No BYOK gymnastics required this time: Cursor added Claude Opus 5 to the native model picker on launch day, July 24 — a pleasant contrast to the two-week lag Opus 4.8 saw in June. Open Settings → Models, enable claude-opus-5, and it’s available in Chat and Agent on paid plans.

If you’d rather meter it yourself, the BYOK route works as always: add your Anthropic API key under Settings → Models → API Keys, and Cursor bills nothing while Anthropic charges you $5/$25 per million. For heavy agent use, the flat Cursor plan is usually the better deal; for occasional Opus-class escalation on top of a mostly-Sonnet workflow, BYOK gives you cost visibility per session.

Using it in Cline and Claude Code

Cline’s Anthropic provider takes a model ID directly, so this works regardless of whether your extension build has updated its dropdown yet: Settings → API Provider → Anthropic, and set the model to claude-opus-5. If the picker doesn’t list it, type the ID into the custom model field — it’s a standard Messages API model with tool use, prompt caching, and the full 1M context, so nothing else changes.

Claude Code got Opus 5 on July 24 as its default Opus model. Verify what you’re running with /model, or pin it explicitly:

claude --model claude-opus-5

Expect one behavioral surprise in agent logs: Opus 5 narrates its progress more than 4.8 did and delegates to subagents more readily. That’s documented behavior, not a misconfiguration.

The effort ladder is the real new control surface

Opus 5 supports all five effort levels — low, medium, high, xhigh, max — via output_config.effort, defaulting to high. On a model this capable, effort is where your cost control actually lives, because it scales everything: thinking depth, tool-call count, and exploration.

response = client.messages.create(
    model="claude-opus-5",
    max_tokens=64000,
    output_config={"effort": "medium"},
    messages=[{"role": "user", "content": "Refactor the auth module to use the new session store."}],
)

Anthropic’s own guidance for Opus 5 differs from the 4.7/4.8 era, where the advice was “start at xhigh for coding.” For Opus 5: start at high (the default), step up to xhigh only for demanding long-horizon agentic work, reserve max for tasks that justify unconstrained token spend, and — their words — use low and medium “liberally” wherever your evals show quality holds. The docs specifically call out efficiency at lower effort as one of the model’s biggest gains, and code review as a task that stays accurate at reduced effort. If you tuned effort settings on 4.8, rerun the sweep rather than carrying them over.

Three traps around this parameter, all verified against the official docs:

1. Changing effort mid-conversation invalidates your prompt cache. Effort shapes the rendered prompt, so flipping from high to low between turns throws away cached prefixes — at 1M-context session sizes, that’s real money. Pick a level per session and hold it.
2. Effort doesn’t shorten responses. On Opus 5 it controls thinking volume, not visible output length. If replies are too long — and Opus 5’s defaults run noticeably longer than 4.8’s — prompt for brevity instead of lowering effort.
3. The breaking change: thinking: {"type": "disabled"} now returns a 400 error at xhigh or max effort. On Opus 4.8, disabling thinking was independent of effort; on Opus 5 it’s only accepted at high or below. Pipelines that pinned thinking off while cranking effort up will fail on migration — keep thinking disabled and drop to high, or keep the effort level and delete the thinking field.

Related quirk: with thinking disabled, Opus 5 can occasionally write a tool call into its text output instead of emitting a proper tool_use block, or leak internal XML tags into visible responses. Anthropic’s recommendation is to leave thinking on and control cost with effort — for coding agents, that’s the right call.

Migrating from Opus 4.8: what else breaks

The model ID swap is the easy part. The rest of the checklist:

• Thinking is on by default now. Opus 4.8 ran without thinking unless you opted in with thinking: {"type": "adaptive"}. Opus 5 thinks by default, and max_tokens is a hard cap on total output — thinking plus response — so workloads that ran thinking-free on 4.8 with tight max_tokens values will now hit the ceiling mid-task. Raise them; 64K is a sane starting point for agentic work.
• Sampling rules carry over. Non-default temperature, top_p, or top_k already returned 400 errors on Opus 4.8, and that holds. Tone control belongs in the system prompt.
• Delete your verification boilerplate. Opus 5 verifies its own work unprompted. Instructions like “include a final verification step” or “spawn a subagent to double-check” — standard armor in 4.8-era prompt kits — now cause over-verification, burning tokens on redundant re-checking. This is the least obvious migration item and the one most likely to inflate your bill invisibly.
• Two new betas worth knowing: mid-conversation tool changes (mid-conversation-tool-changes-2026-07-01 header) let you add or remove tools between turns without nuking the prompt cache — a genuine win for long MCP-heavy sessions — and the fallbacks parameter gains a "default" mode (server-side-fallback-2026-07-01) that applies Anthropic’s recommended fallback routing without you maintaining a model list.

The cost math

A typical Cursor or Cline agent session — call it 50K input tokens, 8K output:

Backend	Input	Output	Per session	50 sessions/day, monthly
Sonnet 5 (intro, to Aug 31)	$0.100	$0.080	$0.18	~$270
Sonnet 5 (from Sep 1)	$0.150	$0.120	$0.27	~$405
Opus 5	$0.250	$0.200	$0.45	~$675
Opus 5 fast mode	$0.500	$0.400	$0.90	~$1,350
Fable 5	$0.500	$0.400	$0.90	~$1,350*

*Fable 5’s per-session sticker matches Opus 5 fast mode, but its always-on deep reasoning typically generates more output tokens per task, so real Fable 5 bills run above this line.

Three readings of that table. First, Opus 5 at $0.45 versus Fable 5 at $0.90 for a 1.1-point SWE-bench Pro gap is the whole launch in one line — Anthropic just cut the price of near-frontier coding in half. Second, from September 1 the Opus 5 premium over Sonnet 5 shrinks to about 1.7x per session, which makes “Opus for anything multi-file, Sonnet for the rest” a much easier routing rule to justify than it was in the 4.8 era. Third, fast mode at double price is for latency-bound work — CI-gating review, live pairing — same model, same answers, roughly 2.5x the output speed, exactly as it works on Opus 4.8’s fast mode.

Prompt caching moves these numbers substantially in long sessions: cache hits bill at $0.50/M, a tenth of base input, and the new 512-token cache minimum means even lean system prompts qualify.

If your API bills are climbing regardless of model choice, the escape hatch is a local fallback for routine work — see runaihome.com’s guide to the best local models by VRAM and the open-weight coding options tracked at aifoss.dev, then reserve Opus 5 for the sessions that earn it.

When Fable 5 is still worth 2x

A verdict that names the loser: Opus 4.8 is done as a daily driver. Same price, ten points worse on SWE-bench Pro, weaker at every effort level — its only remaining role is as a pinned fallback for pipelines that can’t absorb Opus 5’s thinking-default changes yet.

Fable 5 survives in a narrower role. Pay double when you’re on genuinely frontier ground: novel architecture design where the 0.8-point SWE-bench Pro gap understates the difference in reasoning ceiling, research-grade analysis, and terminal-heavy agentic work where Fable 5’s verified 88.0% Terminal-Bench 2.1 score stands unanswered — Opus 5 hasn’t posted a number there yet. If you’re on a Claude subscription rather than the API, the calculus tilts further toward Opus 5: it’s included as the Max default and Pro’s top model, while Fable 5 access remains capped and credit-metered with its classifier-reroute quirks.

For everyone else — the developer running Sonnet 5 daily and escalating hard tasks upward, the team lead pricing out backends in our cost comparison, the BYOK tinkerer who just wired up Kimi K3 or Gemini 3.6 Flash — the escalation target just got twice as cheap. Route accordingly.

FAQ

Is Claude Opus 5 available in Cursor right now? Yes — it was added to Cursor’s native model picker on July 24, 2026, launch day. Enable it under Settings → Models. Cursor’s own testing has it matching Fable 5 on CursorBench (66.7% vs 66.5%) at half the cost per task.

What’s the exact model ID for the API, Cline, and Claude Code? claude-opus-5 on the Claude API, Vertex AI, and in Cline/Claude Code configs; anthropic.claude-opus-5 on Amazon Bedrock. It’s a pinned snapshot despite having no date suffix.

Did the price go up from Opus 4.8? No. $5/$25 per million tokens, unchanged. Batch is $2.50/$12.50, cache hits $0.50/M, and the optional fast mode is $10/$50 on the Claude API only.

Will my Opus 4.8 code break if I just swap the model ID? Two things to check: thinking is now on by default (so max_tokens must cover thinking plus response), and thinking: {"type": "disabled"} combined with xhigh or max effort returns a 400 error. Also remove any “verify your work” prompt instructions — Opus 5 self-verifies and will double-spend on them.

Is the 96% SWE-bench Verified number real? It’s Anthropic’s system-card figure (five-trial average), and the independent vals.ai leaderboard currently shows 97.0% — so yes, though SWE-bench Verified is effectively saturated at the frontier. The more informative number is 79.2% on SWE-bench Pro, where Fable 5 still leads by 0.8 points.

Should I use fast mode? Only when latency blocks something — a CI gate, a live pairing session. It’s the same model at ~2.5x output speed for 2x price, in research preview, API-only.

Sources

• Introducing Claude Opus 5 — Anthropic
• Model pricing — Claude Platform Docs
• Models overview — Claude Platform Docs
• What’s new in Claude Opus 5 — Claude Platform Docs
• Effort parameter — Claude Platform Docs
• Claude Opus 5 now available — Cursor Community Forum
• SWE-bench Verified leaderboard — vals.ai
• Claude Opus 5 Benchmarks Explained — Vellum
• Meet the New Claude Opus 5 — MarkTechPost
• Anthropic upgrades Claude with new Opus 5 model — 9to5Mac

Last verified July 27, 2026. Pricing and model availability change frequently; check the official pricing page before committing to a backend.