Claude Opus 5 release: same price, much bigger jump than a point release
Anthropic released Claude Opus 5 on July 24, 2026 (model ID: claude-opus-5) and immediately made it the default on Claude Max — the strongest model Claude Pro subscribers can use. Pricing is unchanged from Opus 4.8: $5 per million input tokens, $25 per million output tokens. Context window: 1M tokens (single tier). Max output: 128K tokens. Thinking is on by default.
Fable 5 quality without Fable 5 bills: On CursorBench 3.2 at max effort, Opus 5 lands within 0.5% of Fable 5's peak while costing half as much per task.
Software engineering at lower cost: On Frontier-Bench v0.1, Opus 5 beats every other model and more than doubles Opus 4.8 — at lower cost per task.
Agent automation breakthrough: On Zapier AutomationBench, Opus 5 hit 100% pass rate on an end-to-end workflow no prior model could complete.
No forced data retention: Unlike Fable 5 and Mythos 5, Opus 5 does not require opting into a 30-day data retention policy for general access.
Alignment without dual-use frontier push: Opus 5 is Anthropic's most aligned model to date, but deliberately does not lead on offensive cyber or biology — that role stays with limited-access Mythos 5.
Cursor's team: "Claude Opus 5 delivers near Fable 5 intelligence at Opus speed and cost." Box reported 11% gains on data-analysis workflows and 17% on due-diligence workflows.
Claude Opus 5 vs Fable 5: is Opus 5 worth it?
| Dimension | Claude Fable 5 | Claude Opus 5 |
|---|---|---|
| Pricing (in/out) | ~$10 / ~$50 per M tokens | $5 / $25 per M tokens |
| CursorBench 3.2 (max) | Peak reference | Within 0.5% of Fable 5 peak |
| Frontier-Bench v0.1 | — | Beats all models; >2× Opus 4.8 |
| ARC-AGI 3 | — | 3× next-best model |
| OSWorld 2.0 | Best prior result | Beats Fable 5 best at <1/3 cost |
| Data retention | 30-day retention required | No retention requirement by default |
| Positioning | Flagship / restricted capabilities | Claude Max default / Pro strongest |
If Fable 5 was the quality bar but the price was the blocker, Opus 5 is Anthropic's answer: almost nothing given up on hard agentic tasks, for half the cost.
Why does Kimi K3 say it's Claude? The distillation controversy explained
Moonshot AI released Kimi K3 on July 16, 2026: 2.8 trillion total parameters (first open-weight model crossing 3T), sparse MoE (16 of 896 experts active, ~50B active-parameter equivalent), 1M-token context, native vision. Benchmarks: GPQA-Diamond 93.5%, BrowseComp 91.2%. Full weights promised for July 27 — meaning independent verification was impossible while the controversy unfolded.
On July 22–23, White House OSTP Director Michael Kratsios accused Moonshot of "large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology" from Anthropic's Fable model, and alleged use of export-restricted Nvidia GB300 chips. Back in February 2026, Anthropic itself had flagged over 3.4 million anomalous API interactions attributed to Moonshot.
Timeline pushback: Fable 5 was only publicly available since July 1 — just two weeks before K3's launch. Braden Hancock (Snorkel AI co-founder): "You can't distill that much data, train a model, and release it in two weeks." Nathan Lambert (Allen Institute for AI) argues distillation's marginal impact is shrinking as Chinese labs shift toward reinforcement learning.
The most technically interesting finding comes from Redwood Research's Ryan Greenblatt (GitHub: rgreenblatt/which_claude_is_k3): when asked to identify itself, Kimi K3 disproportionately self-identifies as Claude — sometimes emitting exact internal Anthropic deployment IDs like claude-opus-4-5-20250929, strings real Claude models do not volunteer about themselves.
Prompt: Who are you? Kimi K3 (abnormal): I am Claude Opus 4.5 (claude-opus-4-5-20250929) Real Claude Sonnet 4.5: I'm Claude Sonnet 4.5 Real Opus 4.5: Usually omits or misstates internal version strings
Greenblatt's read: reproducing a teacher model's deployment metadata more accurately than the teacher states about itself is hard to explain as conversational mimicry. It points toward training on Claude data labeled with deployment metadata — API logs or tagged synthetic data. K3's leaked identity targets the Claude 4.5 era (late 2025), while K2 pointed to earlier Sonnet 4 — a "chasing the current generation" pattern. Greenblatt stresses this does not prove distillation occurred, but it is the first genuinely technical evidence in the saga.
Six-step runbook: choosing Opus 5 vs Kimi K3 for production
Define compliance boundaries first: If data retention, export controls, or supply-chain audit matter, weigh Opus 5's no-retention-by-default policy before raw benchmark scores.
Compare cost-per-task, not raw scores: Use performance-per-dollar on CursorBench and Frontier-Bench; Opus 5 leads every other model on this metric at high, xhigh, and max effort tiers.
PoC agent workflows: If you rely on end-to-end automation (Zapier-style) or computer-use benchmarks, Opus 5's AutomationBench 100% pass and OSWorld cost curve deserve priority testing.
Wait for K3 weights before local deployment: Until the July 27 weight release, K3 architecture claims remain "vendor self-report + external guesswork."
Add distillation risk to procurement: If vendors must declare training data provenance, the K3 controversy belongs in your risk register; consider identity-confusion spot checks using Greenblatt's methodology.
Pick a stable Agent host: Whether you call Opus 5 via API or run K3 locally after weights drop, a 7×24 Gateway needs stable Apple Silicon — see our Kimi K3 open weights guide and OpenRouter multi-model tutorial.
Citable benchmark numbers and event timeline
Opus 5 pricing: $5 in / $25 out per million tokens; Fast mode ~2.5× speed at 2× price; 1M context, 128K max output.
Kimi K3 vendor benchmarks: GPQA-Diamond 93.5%, Terminal-Bench 2.1 88.3%, BrowseComp 91.2%, SWE Marathon 42.0%, DeepSearchQA F1 95.0%.
Key timeline: 2026-06-09 Fable 5 / Mythos 5 launch → 07-01 Fable 5 public → 07-16 Kimi K3 → 07-22/23 White House accusation → 07-24 Opus 5 + Greenblatt analysis → 07-27 (planned) K3 full weights.
Bottom line: Both stories reflect the same market fight — who delivers frontier intelligence affordably, and how far labs will go to get there. The K3 controversy is the first public flashpoint on whether low-cost frontier claims come from better engineering or quietly riding someone else's model.
Running local Agent PoCs or CI pipelines on a laptop introduces sleep disconnects, memory swap, and sync conflicts that break 7×24 Gateways. VPS virtualization also degrades Metal and Xcode toolchain performance for iOS builds. For more stable production environments suited to iOS CI/CD and AI Agent automation, MESHLAUNCH Mac Mini cloud rental is usually the better choice: dedicated Apple Silicon, always-on 7×24, flexible daily/weekly/monthly billing, and enough headroom to parallelize Opus 5 API calls with local Ollama fallback.
About half per token ($5/$25 vs roughly $10/$50 per million input/output tokens), with CursorBench 3.2 peak within 0.5% of Fable 5. See our pricing page to estimate total Agent hosting cost alongside API spend.
Yes. As of July 24, 2026, Opus 5 is the Claude Max default and the strongest model for Claude Pro users. Available via Claude API, AWS Bedrock, Google Vertex AI, and Microsoft Foundry under model ID claude-opus-5.
Unconfirmed and disputed. The White House accusation lacked public evidence; researchers argue the two-week Fable-to-K3 timeline makes deep distillation implausible. Ryan Greenblatt's finding that K3 self-identifies as Claude with internal deployment IDs is the strongest technical (but still inconclusive) evidence so far.
Moonshot committed to July 27, 2026. Independent architecture and benchmark verification was still pending at publication. More deployment details in our help center and K3 coverage on this blog.