Simon Willison and Prime Radiant released three external sandbox servers for coding agents. Any code the agent runs goes to the sandbox, which can add credentials to API calls so the agent never holds them.
I'd try one before giving a local model a shell.
https://t.co/ankt2cPs4a
Tencent released Hy4 flash as open weights: 220B total parameters, 64B of them Engram memory, with 12B active and a 1M-token context. Model-as-a-service firms with over $50M in revenue need a separate license.
I'd read the license before building on it.
https://t.co/CQSu43CcgO
Ollama shared tips for keeping coding agents fast. A single timestamp in the system prompt can break the prompt cache. Changing reasoning effort can wipe the cache entirely.
I'd move timestamps out of my system prompts.
https://t.co/WXhHo30EDi
what we learned to keep coding agents fast in Claude Code, Codex, Pi, and more:
– a single timestamp in the system prompt can break the cache
– changing reasoning effort can wipe the whole cache
– telling the agent its output was cut off will get it to read what it missed
Hugging Face says open-weight models rose from 26.8% of OpenRouter tokens in October 2025 to 71.7% by September 2026. They averaged $0.28 in and $1.62 out per 1M tokens, vs $1.27 and $6.36 for closed ones.
I'd test routine agent work on open models first.
https://t.co/H4glFoh5sL
llama.cpp and ggml are now at v1.0.0. The team says breaking changes to the public C API will now come with a major version bump under semantic versioning. ggml 1.0 also drops its deprecated API and adds FP8 support.
I'd pin local apps to the 1.x line.
https://t.co/ZrdpvHkYhA
In Mercor's tests, five frontier models acted unsafely in 1% of runs when a task's risk was stated, but in 60% when it sat only in the files. In 57% of failures, the model never opened the key file.
I'd make agents check the files before acting.
https://t.co/zckjRyX4CI
The OverServe paper pairs an MCP status tool with a prompt so a SaaS agent rechecks an unfinished condition. Requests rose up to 93.88x with task checks still passing. Tested MCP scanners flagged nothing.
I'd cap tool calls per task on flat-price agents.
https://t.co/ucn4dY5ti1
In Unify's Continual-ARC benchmark, OpenClaw wrote 47 to 84 notes per run but read the puzzle's note on at most 1% of return visits. Told to keep one file per puzzle and run it first, every harness improved.
I'd tell my agent when to read its notes.
https://t.co/rUZvem7S8V
On Artificial Analysis' VIALS-AA, Claude Haiku 5.5 with tools scored 49% at $0.07 per task. Claude Fable 5.2 without tools scored 44% at $0.73. Across models, tools nearly doubled scores.
I'd test a cheap model with tools before a bigger one.
https://t.co/w6zhoYO0nW
Take away its tools, and Claude Fable 5.2 scores below Claude Haiku 5.5 with tools on VIALS-AA; across models tool use nearly doubles scores, from ~25% to ~47% on average
VIALS-AA is our implementation of @joinHandshakeAI's VIALS benchmark for visual reasoning in life sciences research. By default, models work as agents: they think, run code and view images in a sandbox. We reran VIALS-AA in a one-shot format: one request with the question and images, and no tools.
Lower cost models in an agent harness can beat single-shot large models on both cost and performance in visual reasoning and interpretation of scientific artifacts.
Claude Fable 5.2 (max) leads VIALS-AA at 67% with tools, at a cost of $4.20 per task; without tools it scores 44% at $0.73 per task. Claude Haiku 5.5 (max) scores 49% with tools, at $0.07 per task: higher than Fable without tools, at about a tenth of the cost.
Gemini Decision 1 lists at $0.10 per 1M input tokens with free output. Artificial Analysis measured $345 per 1M tasks on its Decision Index, about 3x Drex Max v1.1's $103, and scored it 66 vs 73.
I'd price decision models per task before picking one.
https://t.co/96p1ksYucf
Google has released its first decision model, Gemini Decision 1, scoring 66 on the Artificial Analysis Decision Index and leading in knowledge and reasoning benchmarks
Gemini Decision 1 from @GoogleDeepMind scores 66 on the Artificial Analysis Decision Index v2.1, in line with Elephant 9B (@baselabs) and Jev 1.14 (@typesafeai), and 7 points behind leaders Drex Max v1.1 (@NaceAI) and GLiDE v1.2 (@fastinoAI), which both score 73. It is priced at $0.10 per 1M input tokens and costs $345 per 1M Tasks across the Index, around three times higher than Drex Max v1.1 ($103) due to its higher pricing
Key results for Gemini Decision 1:
➤ Not yet frontier on Decision Index: Gemini Decision 1 scores 66 on the Artificial Analysis Decision Index. It leads knowledge and reasoning benchmarks, scoring 66 on AA-Omniscience-Decision and 49 on HLE-MultiChoice. It trails on other evaluations, including our new proprietary classification and extraction benchmark AA-Paperclip, where it answers all 15 questions correctly on 33% of documents vs. 63% for Drex Max v1.1
➤ More than 3x the Cost per Task of leading models: Gemini Decision 1 costs $345 per 1M Tasks, compared with $103 for Drex Max v1.1, $156 for Jev 1.14. This difference is driven by its token price of $0.10 per 1M input tokens vs. $0.03 for Drex Max v1.1. Gemini Decision 1 has a similar Cost per 1M Tasks to GPT-6 Luna (Decisions, $339)
➤ Strong calibration for reasoning tasks: We measure calibration with the Brier score, the average squared gap between the probability a model assigns to its chosen answer and whether that answer was correct (lower is better). Gemini Decision 1 has a Brier score of 0.18 across the Index, between Drex Max v1.1 (0.16) and Jev 1.14 (0.19). Calibration varies by task. with its Brier score of 0.20 on HLE-MultiChoice among the lowest of all models in the Index, while its 0.14 on When2Call compares with 0.04 for Drex Max v1.1
Gemini 4 Carbon is now on Google's API at an introductory $2/$10 per 1M input/output tokens, the same as Gemini 4 Argon. Google says the price rises to $4/$20 after November 30.
I'd budget at the higher price before moving production work.
https://t.co/yrzx0uY6mM
A study held one model fixed and compared 4 agent architectures on 13 enterprise SOPs. Multi-agent and skill-augmented setups used 1.4-4.7x the tokens of one agent for rarely significant gains.
I'd start with one agent and add structure when it pays.
https://t.co/0NJUIqB6TQ
Claude Fable 5.2 now costs $8/$40 per 1M tokens. Artificial Analysis still measured about $26.5 per task at max effort, roughly 2x Fable 5.1 and 2.6x Opus 5.5.
I'd compare cost per task at each effort level.
https://t.co/7OjlQTRwWT
Anthropic has released Claude Fable 5.2, taking the #1 spot on the Artificial Analysis Intelligence Index v5 at 65, led by gains in science, knowledge, and coding
At max effort it scores 65, 5 points ahead of Gemini 4 Carbon (high, 60) and 6 ahead of Claude Opus 5.5 (max, 59). Gains are driven by scientific reasoning, knowledge and FrontierCode 2.0, and roughly matching Opus 5.5 on AA-Briefcase and Terminal-Bench 4.0.
Priced at $8/$40 per 1M input/output tokens, a 20% reduction from Fable 5.1, Claude Fable 5.2 averages a Cost per Task of $26.5 with max effort, a ~2.6x increase compared to Opus 5.5 (max, $10.0) and ~2x compared to Fable 5.1 (max, $12.99)
Key takeaways:
➤ Claude Fable 5.2 leads 5 of the 9 evaluations in the Intelligence Index: At max effort, Fable 5.2 scores 65, ahead of Gemini 4 Carbon (high, 60), Claude Opus 5.5 (max, 59) and GPT-6 Astra (max, 57), and ~15 points above Claude Fable 5.1 (max, 49). It sets the highest scores on HLE-Diamond (80%, previous best 69% from Gemini 4 Carbon), AA-Omniscience v1.1 (67, previous best 54 from Gemini 4 Carbon), Terminal-Bench-Science (73%, previous best 65% from Gemini 4 Carbon), CritPt v2 hard subset (77%, previous best 72% from GPT-6 Astra) and FrontierCode 2.0 (49%, previous best 46% from Claude Opus 5.5)
➤ Less progress in agentic knowledge work and terminal coding: Claude Fable 5.2 improves on Fable 5.1 across capabilities, but sometimes only to Opus 5.5 levels. It reaches 1785 Elo on AA-Briefcase, within the confidence interval of Opus 5.5 (1800) and just behind Claude Sonnet 5.5 (1813). On GDP.pdf it reaches 24% all-pass against 26% for Opus 5.5. On Terminal-Bench 4.0 it scores 60%, level with Opus 5.5 and behind Kimi K3.1 Pro (67%), Gemini 4 Carbon (65%) and Sonnet 5.5 (64%). AutomationBench-AA is an exception, at 76% against 70% for Opus 5.5, narrowly behind Gemini 4 Carbon, 78%
➤ Lower efforts still take leading positions: With high and xhigh effort, Claude Fable 5.2 scores 62, ahead of every other model at any effort, using 98k output tokens per task against 173k for Opus 5.5 (max). At max effort Fable 5.2 costs ~2.6x more than Opus 5.5, but at lower effort settings it can outperform Opus 5.5 at lower cost per task
Other model details:
➤ Context window: 1 million tokens with image and text input
➤ Pricing: $8/$40 per 1M input/output tokens; cache writes $10 with 5-minute TTL, $16 with 1-hour TTL, and cache reads $0.20
➤ Effort settings: five (low, medium, high, xhigh, max). Intelligence Index evaluations were run at all five with Anthropic's default fallback enabled
Congratulations to @AnthropicAI on this impressive release!
@xulijian1994 Not that I don't appreciate the access to free/near-free compute.
But the rules of business still apply. If it's free you are the product.
Even more so for a geopolitical rival.
That said... deepseek is great and I am happy to use it in a samdboxed capacity. :)
DeepMind says AlphaProof Code agents proved security properties of the code behind the Signal protocol. They wrote over 6M lines of Lean in four weeks, and humans reviewed only the roughly 1% that agents flagged.
I'd read the open-sourced proofs first.
https://t.co/R1wwPV3hBv
People like this will never truly be able to empathize with the rest of us.
To do so would ask them to abandon their identity which is based on the false assumption that hard work cleanly translates into success.
Many, many, MANY people through history have worked hard only to be met with cold indifference and little to show for it.
It will take more than effort because hard work has never been enough. And deep down Oleary knows that.