The Intake — Wednesday, September 23, 2026

If you've been running Opus 5 in production and watching cache-read costs accumulate, Anthropic released Opus 5.5 on September 22 with a 60% reduction at that line: from $0.50 to $0.20 per million tokens cached.

Output pricing drops from $25 to $20 per million tokens. Input pricing is $4 per million tokens. Output speed is 30% faster than Opus 5. On Terminal-Bench 4.0, the model scores 66.4%. Fable 5.1 scored 55.8% on the same evaluation. METR and Frontier Design conducted external safety evaluations. US NIST CAISI ran an additional evaluation. Anti-distillation protections are included.

This release also expands the Life Sciences Verification Program for vetted organizations. The Cyber Verification Program also expands. Anthropic states that enrolled users can access capabilities in biology and cybersecurity domains. It says those capabilities match Mythos 5.1. Sonnet 5.5 is expected in the coming weeks. Haiku 5.5 follows on the same timeline. If you're on Opus 5 in production, the output speed gain and cost cuts take effect on any existing workload — no architecture change required.

If you've been routing agent workloads on GPT-5.6 and watching API costs, OpenAI released two successor models on September 22 priced more than 50% below GPT-5.6 promotional pricing.

Sol is positioned for complex tasks and coding. Input pricing is $2 per million tokens. Output pricing is $10 per million tokens. It scores 68.8% on DeepSWE v1.1. OpenAI states it makes half as many mistakes as its predecessor. Luna is positioned for high-volume clerical work. Input pricing is $0.10 per million tokens. Output pricing is $0.50 per million tokens. Luna is available on OpenAI's Free and Pro tiers.

Both models are available in GitHub Copilot across VS Code, JetBrains, and Xcode. Eclipse and GitHub web are also supported. If your pipeline handles both complex and high-volume routine tasks, Sol and Luna span a 20x cost ratio on the input side — a routing option that didn't exist on this platform before September 22.

The UN System Data Commons launched on September 17, 2026, at data.un.org. It is a joint Google and United Nations platform built on Google Data Commons. MCP-standard support for agent queries is included.

Twenty-six UN entities committed to the platform. Approximately twenty are participating at launch. The stated target is 80% of UN statistical datasets accessible through the platform by 2027. Google.org contributed $2 million to the effort. The funding was channeled through the UN Foundation. A UNICEF benchmark tested six major large language models against development indicator queries. Average accuracy was 21.2% — the baseline figure used to motivate the platform. Per the joint announcement, ChatGPT referrals to UNICEF's data site increased 67% year-over-year. The increase was in the period preceding the launch.

If you're building agents that query development or geopolitical statistics, data.un.org is now the primary-source MCP-accessible endpoint for that data.

If you're evaluating frontier models for engineering-domain tasks, the SpaceX training supplement in Grok 4.7 is the disclosed provenance difference in this release.

xAI released Grok 4.7 on September 21, 2026. The model has 2.1 trillion parameters. That is a 40% increase from Grok 4.6's 1.5 trillion. Pricing is unchanged from Grok 4.6. Input is $2 per million tokens, output is $6. On CursorBench 4.0 it scores 46.3%. Grok 4.6 scored 40.4% on the same evaluation. DeepSWE v1.1 score is 71.0% at high effort. EEBench score is 64.0%. GDPval is 1,695, against 1,735 for Claude Fable 5.1. Training data includes a supplemental set from SpaceX. That set covers satellite telemetry and manufacturing records. Engineering failure logs are also included. xAI states the model includes a new safeguard stack. The model is available on the xAI API, Grok app, and Cursor. Grok Build and GitHub Copilot are also supported. xAI's roadmap lists Grok 4.8 and 4.9 as subsequent releases. Grok 5 follows.

If your domain work involves satellite operations, manufacturing, or engineering failure analysis, Grok 4.7's SpaceX supplement is the disclosed training-data advantage specific to those domains.

---