Saturday, July 25

The signal

1. Anthropic ships Claude Opus 5 at half the price of Fable 5 What happened: Anthropic released Claude Opus 5 on July 24; priced at $5/$25 per million input/output tokens (same as Opus 4.8, half of Fable 5’s $10/$50); new effort-toggle (low/medium/high) exposes per-request cost-capability tradeoffs; available immediately on API, Bedrock, Google Cloud, Microsoft Foundry, Claude.ai, and Claude Code; now the default on Claude Max and strongest on Claude Pro. Why it matters: The effort-toggle is the first per-call quality/cost dial on any frontier API; combined with the pricing, it changes the cost math for agentic loops doing mixed-complexity work. Confidence: Confirmed Sources: Anthropic (primary) · Fortune (independent) · Bloomberg (independent)

2. NVIDIA leads 25-company open-weights letter to Congress; OpenAI and Anthropic absent What happened: jensenhuang used his first-ever X post on July 24 to cosign “Open Weights and American AI Leadership” alongside Meta, Microsoft, Palantir, and 21 other organizations, directly targeting the AI Kill Switch Act introduced July 23; elonmusk endorsed the letter the same day; OpenAI and Anthropic did not sign. Why it matters: The open/closed split is now a lobbying fact with structural stakes: NVIDIA’s compute revenue depends on open training runs, giving the 25-company bloc a financial incentive to fight the Kill Switch Act before it reaches the Senate – and they have far more political surface area than the two absent labs. Confidence: Confirmed Sources: Fortune (independent) · Decrypt (independent) · Seeking Alpha (independent)

Since last brief

• natolambert published “Kimi K3: The open-weights escalation” on Interconnects: argues the open-to-closed frontier gap has moved from “6-9 months” to “3-5 months”; calls K3 the strongest open model released, sitting #3 on Artificial Analysis Intelligence Index and #1 in Frontend Code Arena interconnects.ai
• Kimi K3 weights still on track for July 27; TechTimes (July 24) flags undisclosed hallucination risk – Moonshot has not published a hallucination benchmark alongside its performance numbers TechTimes
• TRAINS framework confirmed as “notify-then-ship with 30-day pre-release window”; OpenAI, Anthropic, Google, Microsoft, and xAI signed; Meta has not; August 1 compliance deadline one week out TechTimes
• ExploitGym/Kill Switch Act delta: 25-company coalition now publicly lobbying against the bill; no Senate action yet

Models & research

Claude Opus 5

Claim: Near-Fable-5 intelligence at $5/$25 per million tokens; beats Opus 4.8 on agentic coding, computer use, and multidisciplinary reasoning; effort-toggle (low/medium/high) gives per-call quality-cost control. Evidence: Anthropic-published benchmark comparisons vs. Opus 4.8; available immediately for API evaluation. No independent third-party eval published at time of this brief. Caveat: “Near Fable 5” is not quantified in the release materials. Effort-toggle behavior under sustained reasoning load is uncharacterized externally. Worth your time: Read – directly affects cost modeling for Opus-class agentic workloads. Anthropic · 9to5Mac

Kimi K3 open-weights escalation (Lambert analysis)

Claim: Open-to-closed gap is now 3-5 months, not 6-9; K3 at 2.8T parameters is the clearest evidence yet; natolambert frames this as a structural shift, not an anomaly. Evidence: Artificial Analysis Intelligence Index rankings; Lambert’s analysis on Interconnects. No community benchmarking on the actual weights yet – that begins July 27. Caveat: The gap estimate is qualitative. Numbers will look different once the community gets the weights and runs independent evals. Worth your time: Read if you are making model selection decisions or planning open-weight deployments. interconnects.ai

Tools worth touching

ego-lite (citrolabs)

• Does: Shared-state browser for humans and AI agents; agents inherit your existing cookies, logins, and bookmarks without credential passing
• Why now: Hit #1 on GitHub trending July 24; team’s own benchmarks show 2.5x faster than Vercel’s agent-browser on complex web automation with fewer tokens; MIT license
• Try if: You’re automating logged-in web workflows with Claude Code or similar agents and want agents to share your existing session state directly
• Skip if: Automation targets unauthenticated pages, or you don’t want agents inheriting your browser session
• Links: github.com/citrolabs/ego-lite · runtimewire

alibaba/open-code-review

• Does: Hybrid deterministic + LLM-agent code review CLI; reads Git diffs, lets the agent read full file contents and search the codebase, outputs line-level comments with built-in security rules (NPE, thread safety, XSS, SQL injection); OpenAI and Anthropic compatible
• Why now: 1,066 GitHub stars today; was Alibaba’s internal code review tool for tens of thousands of developers, now MIT-licensed and open
• Try if: You want automated code review that goes deeper than diff-surface feedback; good complement to Claude Code for pre-merge checks
• Skip if: You need ready-made GitHub Actions integration – check the repo for CI workflow status before committing
• Links: github.com/alibaba/open-code-review

Moves on the board

Watchlist

• Kimi K3 open weights drop (expected 2026-07-27): two days out; watch final license terms and hallucination benchmarks when weights hit Hugging Face; community MXFP4 inference economics still unknown (opened 2026-07-23)
• OpenAI ExploitGym fallout: Kill Switch Act in House with 25-company opposition now formed; no Senate vote scheduled, no NIST guidance, AISI cross-lab investigation pending (opened 2026-07-23)
• TRAINS August 1 compliance deadline: one week out; Meta still absent from signatory list; first real enforcement test of the “notify-then-ship” framework (opened 2026-07-25)