Anthropic gives agents a hardware standard, and Fable 5.1 gets "cheaper" without a price cut

The week on the beat

1. Anthropic previewed the Model Hardware Standard, an MCP-style spec for physical devices

MHS is a shared driver specification that lets agents discover and operate lab instruments and robotic arms, model-agnostic and reachable over MCP, with safety limits enforced in the driver rather than in the prompt. Carnegie Mellon went from raw equipment to a finished dose-response curve in eight hours, and QuEra's laser relock improved from 58% to 99.3% across 700 trials. Integration work that took weeks now takes hours, though Anthropic notes Claude still struggles with physical cause and effect, so human oversight stays mandatory.

Read the story →

2. OpenAI rated Astra "Critical" for cyber capability, the first model to hit that threshold

The advanced cybersecurity features will go to selected partners rather than shipping generally, after OpenAI delayed Astra's development to shore up safety work following July's Hugging Face incident. If your roadmap assumed frontier security tooling would arrive as a normal API product, plan for gated access instead.

Read the story →

3. Gemini 3.8 Flash matches Claude Opus 5 on some agentic coding benchmarks, at a rate with an expiry date

The $0.75 and $3.75 per million is introductory pricing running through December 31, and the model spends roughly 30 percent more output tokens per task than 3.7 Flash because it reasons more and calls tools iteratively. Identical headline rate, larger real bill. It is Google's third Flash release in six weeks while the Pro line stays quiet.

Read the story →

4. Fable 5.1 is cheaper without a list price cut

Base pricing holds at $10 and $50 per million, exactly where Fable 5 sat. The savings come from cache reads dropping 75 percent to $0.25 per million and from fewer tokens per task, which is where the "up to 45 percent less for agentic work" number originates. It scores 52.6% on Terminal-Bench-Science against 24.7% for Fable 5, and three API changes are breaking, including the removal of forced tool use, so read the migration notes before bumping a version string.

Read the story →

5. The Justice Department filed in support of OpenAI in the New York Times copyright case

The brief argues that a competitive domestic AI sector is a national priority requiring flexible standards for training data. However the court rules, this is the precedent that settles what the models you build on were allowed to learn from.

Read the story →

6. OpenClaw 2.0 shipped, with a security caveat written into its own docs

Setup now reuses existing API keys, subscriptions and local models, the rebuilt control UI cut test-harness startup from about 1.6 seconds to 575ms, and shared cloud sessions add real multiplayer. The docs state plainly that those shared sessions are not a security boundary, and The Register is considerably blunter about what that means in practice.

Read the story →

Model moves

• New: Claude Fable 5.1 (Anthropic), $10/$50 per 1M, 1M context. Base rate unchanged from Fable 5; the savings sit in cache reads, down 75 percent to $0.25 per 1M. Model page →
• New: Gemini 3.8 Flash (Google DeepMind), $0.75/$3.75 per 1M introductory through December 31, 1M context. Our registry records the standard rate at double that. Model page →
• New: Muse Spark 1.3 (Meta), $1.25/$4.25 per 1M, 1M context. A Contributor tier landed the same day at $0.10/$0.20. Model page →
• New: Hy4 preview (Tencent), $0.834/$2.501 per 1M, 1M context. Tencent claims it beats Z.AI and Moonshot on internal tests. Model page →
• New: Granite 4.2 8B (IBM), $0.10/$0.15 per 1M, 131K context. Model page →
• New: Mercury 2.5 Preview (Inception), list price $0.20/$0.75 per 1M, 260K context. OpenRouter is quoting $0.04/$0.15, which is an 80 percent launch discount rather than the price. Model page →
• Price change: DeepSeek V4 Flash now lists at $0.22/$0.66 per 1M off peak, doubling to $0.44/$1.32 during peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC, weekdays), confirmed on DeepSeek's own docs. Third-party serving sits far below the vendor: the cheapest credible provider we track is at $0.03/$0.16. Model page →
• Spec change: GLM-5.3's context window grew from 1,048,576 to 1,310,720 tokens. Model page →
• Benchmark revision: Epoch re-scored SimpleQA Verified across 23 tracked models on August 28. Scores moved both ways: Claude Haiku 4.5 from 5.9% to 13.2%, Claude Opus 5 from 56.7% to 59.9%, Qwen3-Max from 67.5% to 48.7%. The models did not change, the suite was re-run. Model page →
• Benchmark revision: gpt-oss-20b's reported scores were re-run on August 27, AIME 2024/2025 from 53.9% to 65.3% and GPQA Diamond from 46.0% to 60.8%. Model page →

Personal take

This week, three different models were offered at lower prices, but none of them reduced their listed price. Fable 5.1 remains at $10 and $50 per million, with the savings coming from a 75% reduction in cache reads. Although Gemini 3.8 Flash maintains the same rate as Gemini 3.7, it uses about 30% more output tokens when thinking, so the same rate ends up costing more. Mercury 2.5's price of $0.04 reflects an 80% launch discount. The per-token rate stopped being a good predictor of your real spending some time ago, yet the labs have realised it is still the figure everybody quotes. When preparing your budget for the next quarter, use tokens per task, your cache hit rate, and how much time the introductory pricing still has left.

Until next Thursday, Anmol