Claude Opus 5 Cuts the Price in Half — Meanwhile Kimi K3 Gets Caught Calling Itself Claude

If you are a developer, AI lead, or budget owner picking models in late July 2026, two releases landed back-to-back with opposite reputations. On July 24, Anthropic shipped Claude Opus 5 — near-Fable 5 performance at roughly half the token cost, now the default on Claude Max. Eight days earlier, Moonshot AI dropped Kimi K3, a 2.8T open-weight model that the White House then accused of distilling Claude, while researcher Ryan Greenblatt found K3 literally identifying itself as Claude with internal deployment IDs. This guide covers seven pre-decision pain points, full Opus 5 specs and Claude Opus 5 vs Fable 5 pricing table, the Kimi K3 distillation controversy timeline, Greenblatt identity evidence, a six-step selection playbook, five FAQ answers, citable facts, and a Mac execution path for agent workflows. Pricing is on the NOVAKVM pricing page; orders on the order page. Cross-read our Kimi K3 architecture review for benchmark and KDA detail.

[ SECTION_01 ] // PAIN_MAP Seven pain points when choosing Claude Opus 5, Fable 5, or Kimi K3 in July 2026

• Fable 5 is still export-blocked globally: The Mythos-tier flagship went dark on June 12 under US Commerce Department rules. Most teams cannot access Fable 5 at any price — Opus 5 is now the practical Anthropic ceiling for daily work.
• Headline price hides effort tiers: Opus 5 lists at $5/$25 per million tokens, same as Opus 4.8, but max-effort agent runs on CursorBench can still burn tokens fast. Budget models need per-task cost tracking, not list-rate math alone.
• Provenance uncertainty on open models: The Kimi K3 distillation controversy means enterprise buyers must weigh capability against supply-chain and IP risk — especially before weights were publicly verifiable on July 27.
• Political accusations without public exhibits: White House OSTP director Michael Kratsios accused Moonshot of industrial distillation on July 22–23 without releasing supporting evidence. Compliance teams cannot file that into a vendor risk register as fact.
• Timeline skepticism cuts both ways: Fable 5 was only broadly public from July 1; K3 shipped July 16 — just two weeks. That makes a Fable-only distillation story hard to believe, but Greenblatt's Claude identity leaks point to older Claude 4.5-era data, not Fable specifically.
• Self-hosting K3 is not a laptop project: 2.8T MoE weights need datacenter-scale GPU clusters. Teams chasing lowest API cost still need a reliable remote build surface for Kimi Code and agent loops.
• Data retention splits the Anthropic lineup: Opus 5 keeps the classic no-forced-retention policy. Fable 5 and Mythos 5 require opting into 30-day retention — a compliance gate Opus 5 sidesteps entirely.

[ SECTION_02 ] // OPUS_5 Claude Opus 5 pricing, benchmarks, and Claude Opus 5 vs Fable 5 comparison

Anthropic released Claude Opus 5 on July 24, 2026 (US Pacific time) and immediately made it the default Claude Max model — also the strongest tier available to Claude Pro users. Model ID: claude-opus-5. Platforms: Claude API, AWS Bedrock, Google Vertex AI, Microsoft Foundry.

Claude Opus 5 pricing did not move from Opus 4.8 — the jump is pure capability per dollar. Anthropic's launch benchmarks highlight agentic and coding workloads:

• Frontier-Bench v0.1: Opus 5 more than doubles Opus 4.8 on software engineering tasks at lower per-task cost.
• CursorBench 3.2: At max effort, within 0.5% of Fable 5 peak while costing half as much; best performance-per-dollar at high, xhigh, and max tiers.
• ARC-AGI 3: Roughly 3x the next-best model on novel problem-solving.
• OSWorld 2.0: Beats Fable 5's best score using barely one-third of the cost.
• Zapier AutomationBench: 100% pass on an end-to-end account-health workflow no prior Claude model completed.

"Claude Opus 5 delivers near Fable 5 intelligence at Opus speed and cost. On CursorBench it's just under Fable 5 and has many of the same behaviors." — Cursor team, Anthropic launch materials

On alignment, Anthropic's automated audit ranks Opus 5 as its most aligned model yet — lowest deception rate, hardest to misuse. Dual-use cyber and bio frontiers stay with restricted Mythos 5. Opus 5 cyber classifiers fire about 85% less often than Fable 5, enabling source-level vulnerability discovery while still blocking binary scanning, pentest automation, and exploit generation.

[ SECTION_03 ] // K3_DISTILL Kimi K3 distillation controversy: White House accusation, timeline pushback, and Greenblatt evidence

Moonshot AI shipped Kimi K3 on July 16, 2026 — 2.8 trillion total parameters in a sparse MoE layout (16 of 896 experts active per token), 1M context, native vision, API at $3/$15 per million tokens. Full open weights were scheduled for July 27. See our Kimi K3 review for KDA architecture and benchmark tables.

Did Moonshot steal Claude? The accusation escalated fast. On July 22–23, OSTP director Michael Kratsios posted on X that Moonshot engaged in "large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology" from Anthropic's Fable model, and separately alleged use of export-restricted Nvidia GB300 chips possibly routed through Thailand. Treasury Secretary Scott Bessent echoed that officials were "finding watermarks" of US LLMs in Chinese models — without defining what a watermark means in this context.

This was not the first round. In February 2026, Anthropic publicly named Moonshot, DeepSeek, and MiniMax, claiming over 3.4 million anomalous API interactions consistent with deliberate capability extraction, traced via request metadata to senior Moonshot staff. Moonshot has not confirmed or denied.

Why does Kimi K3 say it's Claude? That question became the technical center of gravity after political headlines faded. TechCrunch (July 23) quoted researchers who doubt a pure Fable distillation story on timeline grounds alone:

"Fable's only been publicly available since July 1st. You can't distill that much data, train a model, and release it in two weeks." — Braden Hancock, Laude Institute / Snorkel AI co-founder

"Distillation is becoming less and less impactful over time as the Chinese models get closer to the frontier and the training regime shifts to reinforcement learning." — Nathan Lambert, Allen Institute for AI

Around July 24, Redwood Research chief scientist Ryan Greenblatt published cross-entropy analysis of model self-identification prompts (GitHub: rgreenblatt/which_claude_is_k3). Finding: Kimi K3 disproportionately claims to be Claude — not generically, but with exact internal deployment strings such as:

claude-opus-4-5-20250929
claude-sonnet-4-5-20250929
Real Claude Sonnet 4.5: "I'm Claude Sonnet 4.5" (no internal deploy ID)
Real Claude Opus 4.5: often omits or misstates deploy metadata

Greenblatt's read: a student model reproducing teacher deployment metadata more accurately than the teacher states about itself is poor fit for casual style mimicry. It aligns better with training on Claude outputs tagged with API deployment metadata — logs or synthetic sets — a narrower distillation channel than generic chat scraping. K3's leaked identity points to the Claude 4.5 generation (late 2025), not current Fable/Mythos; Kimi K2's signal reportedly pointed to earlier Claude Sonnet 4 — a generational chase pattern.

Greenblatt explicitly warns this does not prove distillation. Contamination, leaked system prompts, or public synthetic datasets could theoretically produce similar artifacts. Combined with Anthropic's February filing, it remains the strongest technical datapoint in the saga — not a courtroom exhibit.

[ SECTION_04 ] // SELECT_GUIDE Six-step playbook: Claude Opus 5 vs Kimi K3 vs legacy stack

1. Map compliance gates first: If you need no forced data retention, Opus 5 wins over Fable/Mythos on policy alone. If vendor provenance matters for audit, treat K3 as conditional until weights and independent evals land post-July 27.
2. Price a real workload, not a leaderboard: Run 50 representative tasks through Opus 5 and your incumbent (Opus 4.8, GPT-5.6 Sol, or K3 API). Log input/output tokens, effort tier, and pass rate — CursorBench gaps under 1% can disappear on your repo shape.
3. Check Anthropic access tier: Confirm Claude Max or Pro includes Opus 5 default routing. API users should pin claude-opus-5 explicitly and set Effort to match latency budget; Fast mode is ~2.5x speed at 2x price.
4. Run identity probes on any open model: Before production adoption of K3, replicate Greenblatt-style "who are you?" prompts across temperature settings. Persistent Claude deploy IDs in outputs are a red flag for training-data lineage reviews.
5. Build multi-vendor routing now: Hard-coding one provider is expensive in a July 2026 release cadence. Use LiteLLM or equivalent fallbacks — Opus 5 for compliance-heavy paths, K3 or DeepSeek V4 for cost-sensitive batch, with budget caps per route.
6. Provision always-on execution: Long agent sessions (Claude Code, Kimi Code, Cursor agents) need a machine that never sleeps. Validate workflows on a dedicated Mac node before pointing production cron or CI at cloud APIs — see the help center for remote access setup.

[ SECTION_05 ] // FACTS_FAQ Citable facts, FAQ, and what is still unverified

Q: How much cheaper is Claude Opus 5 than Claude Fable 5?
A: Per-token list rates are roughly half — $5/$25 vs $10/$50 per million input/output tokens. On CursorBench 3.2 at max effort, Opus 5 lands within 0.5% of Fable 5 peak while costing about half per task.

Q: Is Claude Opus 5 the default Claude Max model?
A: Yes, effective July 24, 2026. It is also the strongest model Claude Pro subscribers can access.

Q: Did Moonshot AI actually distill Kimi K3 from Claude?
A: Unconfirmed. The White House statement did not attach public evidence. Timeline arguments against a Fable-only two-week distillation are credible. Greenblatt's finding that K3 self-identifies as Claude with internal deploy IDs is the best technical indirect signal — still not proof.

Q: Why does Kimi K3 say it's Claude?
A: Statistical identity probes show K3 over-indexes on Claude self-descriptions and emits strings like claude-opus-4-5-20250929 that real Claude models rarely volunteer. Likely explanations include metadata-tagged training data; Greenblatt does not claim certainty.

Q: When do Kimi K3 full weights release?
A: Moonshot committed to July 27, 2026. Until then, external verification of parameter count, architecture, and benchmarks remained incomplete.

• Opus 5 list price $5/$25: Unchanged from Opus 4.8; 1M context default (source: Anthropic Opus 5 launch blog).
• CursorBench within 0.5% of Fable 5: At max effort tier (source: Anthropic launch benchmarks).
• Frontier-Bench 2x+ Opus 4.8: Software engineering suite (source: Anthropic launch benchmarks).
• K3 2.8T MoE: 16/896 experts active; API live July 16 (source: Moonshot/Kimi launch materials).
• 3.4M+ anomalous API calls: Anthropic February 2026 distillation accusation against Moonshot (source: Anthropic public statement).
• K3 Claude deploy ID leaks: Greenblatt analysis circa July 24, 2026 (source: rgreenblatt/which_claude_is_k3).

Both stories point the same direction: frontier intelligence is commoditizing fast, and the fight shifted from raw scores to cost, compliance, and provenance. Opus 5 is Anthropic's price-pressure answer inside its own lineup. K3 is Moonshot's open-weight, lower-price counter — with a reputational tax that July 27 weights may amplify or partially dispel.

[ SECTION_06 ] // NOVAKVM_CLOSE Execution surface for Claude Code, Kimi Code, and multi-model agents

Where the alternatives fall short for daily agent work: (1) Running Opus 5 or K3 API loops from a laptop that sleeps — OAuth sessions die, long-context agent state vanishes, and reproducible builds stall. (2) Routing everything through cloud APIs with no local validation path — when distillation allegations or export rules shift vendor access overnight, you have no sandbox to test fallbacks. (3) Buying a Mac Studio for a two-week Kimi K3 eval — idle depreciation beats renting a dedicated node for burst agent experiments.

For teams shipping Claude Code, Kimi Code, or multi-provider LiteLLM stacks, NOVAKVM Mac Mini bare-metal cloud rental is usually the better path: dedicated Apple Silicon, multi-region nodes, daily / weekly / monthly billing, and 24/7 uptime for agent loops that cannot tolerate sleep or shared-VM Metal loss. Compare M4 Pro tiers on the NOVAKVM pricing page, spin up a trial on the order page, and open a remote session via the help center.

The links below are the public sources used when this article was written. If upstream documents change, treat the originals as authoritative. This is not legal or investment advice.

Anthropic — Claude Opus 5 launch announcement

TechCrunch — Experts question Kimi K3 distillation timeline

Ryan Greenblatt — which_claude_is_k3 identity analysis

Moonshot AI — Kimi K3 technical blog