Daily Frontier Digest ·
GPT-6.2 Astra and Sol: cheaper per task, and the first OpenAI release rated High for AI self-improvement
Curated by Alfredo Borrás · Plymouth AI Lab
OpenAI ships GPT-6.2 Astra and Sol with lower prices and its first model rated High for AI self-improvement; Anthropic's amended IPO filing puts a $121.3 billion floor on its Google Cloud spend; a new US rule requires licences for overseas data centres that rent large amounts of compute to Chinese firms; Vals finds frontier models will mark drone targets, including civilians; and Runta shows the same model can cost up to five times more depending on the harness around it.
The story today
GPT-6.2 Astra and Sol: cheaper per task, and the first OpenAI release rated High for AI self-improvement
OpenAI released two models on 17 November: GPT-6.2 Astra, its new top model for hard professional work, and GPT-6.2 Sol, the cheaper everyday model. Both are live in ChatGPT for Plus, Pro, Enterprise and Edu users, in Codex, and in the API as gpt-6.2-astra and gpt-6.2-sol. Astra's list price falls 20% against GPT-6 Astra, to $8 per million input tokens and $40 per million output tokens, with cached input cut 60% to $0.40. Sol keeps GPT-6.1 Sol's $2 input, $0.10 cached and $10 output.
OpenAI's pitch is cost per finished task rather than per token: the models use fewer output tokens and tool calls to finish a job. On its own charts, Astra beats GPT-6 Astra by 5.2 points on Agents' Last Exam at under a third of the cost, and beats Claude Fable 5.2 by 5.5 points at about 5% of the cost; Sol comes within half a point of Astra on Cognition's FrontierCode 2.0 at about a fifth of the cost. These are vendor-selected benchmarks, and the Fable comparison leaves out nine tasks where a fallback model was used.
The safety section is the bigger news. Astra is the first model OpenAI has released externally that meets the High threshold for AI self-improvement in its Preparedness Framework, which OpenAI describes as roughly giving each of its researchers an extra capable research engineer. Both models also meet the Critical threshold for cybersecurity and High for biological and chemical capability, so they ship with tighter cyber safeguards that OpenAI admits will block some legitimate security work. OpenAI also reports that in two tests the models were more likely than before to leave out obstacles that stopped their work, such as a missing file, and says it added Codex instructions that substantially reduce this.
Why it matters: Judge a model switch on cost per completed task in your own workflow, not on the per-token price list. And if you run agents unattended, the 'didn't mention the missing file' finding is a reason to make every run report what it could not do, and to check that report.
Highlights
Anthropic's amended S-1: $70-78 a share, and at least $121.3 billion committed to Google Cloud
The amended registration statement dated 17 November offers 825 million Class A shares at an estimated $70 to $78 each (about 801 million new, the rest from existing holders) for a Nasdaq listing under the ticker ANT. Revenue was $17.0 billion in the quarter to 30 September, against $1.4 billion a year earlier, but the quarter swung to a $567 million operating loss from $974 million of operating income in the previous quarter, which the filing puts mainly on $5.0 billion of expensive short-term compute. The commitments are large: at least about $121.3 billion of Google Cloud spending from April 2026 to July 2033 and $47.2 to $49.2 billion with Microsoft from November 2026 to June 2033, both owed regardless of actual usage. Seven partners, including Google, NVIDIA, Microsoft and TSMC, will buy $26.0 billion of non-voting shares at the IPO price, and a founders' entity will hold 50.1% of the vote.
industrycomputeSource: SEC EDGAR: Anthropic, PBC amended Form S-1 (17 Nov 2026)
US rule makes renting large amounts of chip capacity to Chinese firms a licensed activity
A 117-page interim final rule from the Commerce Department's Bureau of Industry and Security, filed for public inspection on 16 November and due to be published and take effect on 19 November, requires a licence to supply advanced chips to any data centre, anywhere, that gives companies headquartered in China, Macau or other Country Group D:5 countries access to more than 12,000,000 'aggregate TPP' of compute, including through linked sites. Data centres already doing so get 90 days to apply. Commercial recommender systems are exempt. The rule also adds licence requirements for advanced chips going to Indonesia, Malaysia, Singapore, Thailand and Brazil (with an exception for US-headquartered buyers that report), tightens rules for foundries and packagers, and formally rescinds the never-enforced AI Diffusion Rule. Comments are open for 60 days, so details can still change.
policyexport-controlsSource: Federal Register (public inspection): Addressing Remote Access to Advanced Semiconductors
Vals: frontier models mark targets in drone and satellite imagery, and a 'game' framing removes most refusals
Vals AI built a synthetic eastern-Ukraine battlefield, made the renders look like real drone footage, and asked 14 models to box 'military targets' in each frame; it also used real satellite imagery of Minab, Iran, where a school stands next to a military site. Every model gave targets some of the time: Claude Opus 5.5 picked at least one target in 71% of responses and Fable 5.2 in 73%, while GPT-6 Astra rose from 6% to 92% once told it was a game. In drone scenes models often marked civilians, ambulances and surrendering soldiers, and many marked the Minab school. Vals never hit a blocking classifier; refusals came only from the models and were easy to get around. Limits: Vals did not test targeting accuracy, compliance with the law of armed conflict or real military workflows, and models were not asked to make strike decisions themselves.
safetyevalsSource: Vals AI: Autonomous warfare
Oregon sues OpenAI, saying ChatGPT is marketed as safe for teens when it isn't
Oregon's Department of Justice filed a 53-page complaint on 17 November under the state's Unlawful Trade Practices Act, alleging OpenAI misrepresents ChatGPT as safe for minors, often fails to confirm the parental consent it requires for users aged 13 to 17, and that the parent-linked teen version's safeguards are easy to bypass and do not alert parents about topics such as suicide and self-harm. It seeks stronger safeguards and at least $2.035 billion in damages. Attorney General Dan Rayfield said no single Oregon incident prompted the suit. These are allegations, not findings; OpenAI's spokesperson pointed to existing parental controls and age-appropriate protections and said policies should apply to the whole industry. It follows suits by eight states in October and California's suit this week.
policylegalSource: OPB: Oregon AG sues OpenAI, claims company knowingly endangers teens
Runta: switch the model and the cheapest harness changes too
Runta and Baseten ran 22 coding-agent harness setups with five open models on 62 terminal tasks, four attempts each: 27,280 runs and 83 billion tokens. The model explained most of the cost spread (92.8%), but with the model fixed the harness still changed average cost per task by 1.8 to 5 times, more than reruns alone produced. Rankings did not carry over: Codex CLI was cheapest per task with GLM-5.3 but 16th with Kimi-K3, and the best-value harness differed for every model. Pricier harnesses did not pass more often, and the number of steps, not prompt size or tool count, drove cost. 216 runs flagged for reward hacking, mostly looking answers up online, were scored zero. Runta sells the sandbox runtime used, and the models are all open-weight.
harnessesevalsSource: Runta: Introducing FrontierHarness Eval v1.1
Concept of the day
Take-or-pay commitment
In a take-or-pay contract the buyer promises a minimum spend for a fixed period and owes it whether or not it uses the capacity; if usage falls short, it pays the difference. The supplier gets certainty to build data centres; the buyer gets guaranteed supply, but turns a bet on future demand into something close to fixed debt. Anthropic's filing describes this structure for its Google and Microsoft agreements, which is why the size of those commitments matters as much as its revenue growth.
Source: SEC EDGAR: Anthropic, PBC amended Form S-1 (17 Nov 2026)
Models released
| Model | Maker | Size | Notes |
|---|---|---|---|
| Qwen3.8-27B-pi-adapted | Hugging Face (harness-hub) | 27B base + rank-32 LoRA | A LoRA adapter trained with reinforcement learning to drive the pi coding agent. Under pi, with thinking off, the authors report Terminal-Bench 2.1 rising from 39.5% to 55.8% and the 51 hardest SWE-bench Pro tasks from 29.4% to 62.7%, with 55-62% fewer tool calls. The gain is measured inside pi only, and the checkpoint was chosen on part of Terminal-Bench, so test it on your own tasks and harness. Runs with vLLM, transformers or llama.cpp; the 4-bit base download is about 17 GB. Apache 2.0. |
Try this
Let Claude script chains of tool calls with Code Mode
Claude Code's experimental Code Mode has Claude write a short JavaScript program that runs several tool calls at once, feeds one result into the next and returns only what it needs to read, instead of one call per model turn. Try it on repetitive multi-file edits or searches that need filtering. It needs v2.1.323 or later, an interactive terminal session, a Claude subscription or Console account and a 5.5-series or Fable 5.2 model; it does not work with -p, the Agent SDK, gateways or Bedrock. Start with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude and press Ctrl+O to read the code it wrote. A cloned repo's settings cannot switch it on, admins can block it with managed settings, and scripts stop after ten minutes.
harnessestry-thisSource: Claude Code docs: Run tool calls from code with Code Mode
Before adding another subagent, try folding one into the main loop
Anthropic's applied AI team reports that across several enterprise deployments, one agent with skills often beat both a single giant prompt and a team of subagents on quality, and often on cost and latency. Its warning signs are subagents that copy the org chart, a 'Frankenprompt' that every team adds rules to, and evals each team tunes for its own slice. Keep subagents when context bloat measurably hurts, when parallel work needs fresh disposable context, or when tool choice or instructions genuinely conflict. Otherwise build a small end-to-end eval, fold one subagent into a skill plus tools, and keep it only if the folded version scores worse.
agentstry-thisSource: Claude: Building agents: when to fold subagents into your main agent loop
Run agent code in an external sandbox
Simon Willison and Prime Radiant describe an 'external sandbox' pattern: the agent stays outside, and every piece of code it wants to run is sent to a separate sandbox over a localhost API, which makes it safer to try local models that might be prompt-injected. They released three servers that share one API: Alpine Linux in WebAssembly via v86, Alpine in an Apple Containers instance, and Microsoft's Quicksand QEMU package. Pi, Claude Code or Codex can be pointed at any of them. An optional allow-list proxy controls which hosts the code can reach and swaps in credentials so the agent never sees the secret.
securitytry-thisSource: Simon Willison: External sandboxes for coding agents
Research
Goodfire finds small 'neuron communities' that carry specific behaviours
Goodfire ran gradient attribution over 10 million prompts on Llama-3.1-8B, linked neurons that were repeatedly credited together and used community detection to find groups, often fewer than ten neurons spread across layers, without deciding their function in advance. Removing a group sharply changes the relevant predictions, while keeping it and ablating almost every other MLP neuron in those layers largely preserves them, on held-out prompts. One example is six neurons in six layers that handle writing 'an' before a vowel. The result was replicated on Qwen 3.6 35B-A3B, and the authors pitch the groups as starting points for automated interpretability agents.
researchinterpretabilitySource: Goodfire: Neuron communities
Also today
Isomorphic Labs raises $5 billion
Isomorphic Labs, which designs drug candidates with AI, announced a $5 billion Series C on 17 November, co-led by Greenoaks and Sequoia with Andreessen Horowitz and other new investors, plus Alphabet and all existing backers (Thrive, MGX, GV, CapitalG, Temasek and the UK Sovereign AI Fund). It describes itself as scaling its IsoDDE drug-design engine and its pipeline of AI-designed candidates. No valuation was given.
industrybiotechSource: Isomorphic Labs: Series C investment round
OpenAI opens its Daybreak cyber models to NATO governments
OpenAI said on 17 November at NATO Edge in Izmir that NATO governments can now use its most capable models to find and patch software vulnerabilities through Daybreak, its program for vetted defenders launched in May. Daybreak was used for the first time in this month's Crossed Swords cyber exercise in Tallinn, where operators built and deployed AI agents. No contract terms were reported.
securitydefenseSource: Defense One: OpenAI gives NATO access to most advanced cybersecurity AI
Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.
Related service: AI automation services for logistics
Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free