Daily Frontier Digest ·
OpenAI's own scorecard: agents now complete about half of its two-to-four-hour research tasks
Curated by Alfredo Borrás · Plymouth AI Lab
OpenAI publishes internal numbers on how much its agents now do of its own research; Anthropic prices its IPO at $83 a share, above the range; Google adds Gemini 4 Neon at $3/$15; OpenAI says its models found up to 14% PPA gains on AMD GPU designs; and Microsoft launches seven in-house MAI models priced well below rivals.
The story today
OpenAI's own scorecard: agents now complete about half of its two-to-four-hour research tasks
OpenAI published a second report on how coding agents are changing its internal AI research, framed as tracking indicators of recursive self-improvement, meaning AI speeding up the building of the next AI. Between March and October, its estimated agent success rate on research tasks that would take a person two to four hours rose from 23% to 46%, and on four-to-eight-hour tasks from 13% to 35%.
The gains are uneven by task. Agents succeeded 79% of the time at monitoring training runs, but only 35% at launching them and 25% at writing inference and serving code, which is why OpenAI calls the frontier 'jagged' and says people remain central to the loop. Spending is climbing fast: at API list prices, the median researcher used about $2,400 of agent tokens a day in October (up 62% on September) and the 90th percentile about $23,000 (up 105%). In a survey, researchers' median answer was that they would give up 30% of team compute, or 40% of team headcount, to keep frontier coding agents for three months.
OpenAI also says documented work on safety and alignment grew 77% from January to September, the largest absolute rise of any research area. Every figure is OpenAI measuring itself with internal estimates, and OpenAI notes a temporary harness change in August and September distorted its token-spend series, so treat the trend as indicative rather than audited.
Why it matters: If you run agents on your own work, track the same two things OpenAI does: success rate by task type, not one overall number, and what each person spends a day. The uneven profile (strong at monitoring, weak at production code) is a useful guide to where to hand work over first.
Highlights
Anthropic prices its IPO at $83, above the $70-78 range
Anthropic, PBC priced 980 million Class A shares at $83.00 on 18 November, about $81.3 billion in total. Of those, 956,351,354 shares are sold by Anthropic itself (roughly $79.4 billion before fees) and about 23.6 million by existing holders, whose proceeds Anthropic does not receive. Trading on the Nasdaq Global Select Market under ANT was set to begin on 19 November, with closing expected on 20 November subject to customary conditions. Underwriters have a 30-day option to buy up to 147 million more shares. The price is above the $70 to $78 range in last week's amended S-1. The company's release says nothing about valuation or order demand.
industryanthropicSource: Anthropic: Anthropic prices initial public offering
Google launches Gemini 4 Neon, a mid-tier model at $3/$15 per million tokens
Google released Gemini 4 Neon on 18 November in the Gemini API, AI Studio, Gemini for business, Antigravity and the Gemini app for Pro and Ultra subscribers. It costs $3 per million input tokens ($0.15 cached) and $15 per million output, with a 1 million-token context window. Google positions it between Carbon (hardest long-horizon work) and Hydrogen (fast everyday tasks), and cites a score of 61 on the Artificial Analysis Intelligence Index at its xhigh setting (third overall), 70.9% on the Vals Index (second), 58.6 on Artificial Analysis' Coding Agent Index (first) and 65.2% on Terminal-Bench 4.1. At its default high setting Google says it costs 32% less per task than Carbon. Most other figures, including its lowest-yet fabrication rate, come from Google's internal tests.
modelsgoogleSource: Google: Introducing Gemini 4 Neon
OpenAI says its models improved AMD's flagship GPU design by up to 14%
OpenAI says that working with AMD, its models rewrote SystemVerilog RTL and found up to 14% improvement in power, performance and area (PPA) against heavily optimised baselines on AMD's flagship GPU IP, which it calls roughly a year of progress in under five months. Some changes went into a recent AMD design sent for manufacturing; AMD also verified single-digit PPA gains on key blocks of its next flagship CPU. On OpenAI's own Jalapeño inference chip, model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit, which OpenAI says were proven in silicon. In a public test on two open-source RISC-V cores (9,881 scored runs), GPT-6.2 Astra scored highest. OpenAI is also working with Synopsys on a model trained to use its chip design tools. AMD's figures are reported by OpenAI.
industrychipsSource: OpenAI: Advancing chip design with OpenAI models
Microsoft launches seven in-house MAI models aimed at price, led by MAI Code 2
At Ignite on 17 November, Microsoft AI released a family of models built to undercut rivals on cost. It says MAI Code 2 reaches 98% of GPT-6.1 Sol's coding quality at about 40% of the price and 96% of Claude Opus 5.5's at about 5%, averaged over four benchmarks, and is available in GitHub Copilot; Microsoft says it is not yet tuned for science research or very long agentic work. The lineup also includes MAI Code 2 Flash, MAI Cyber 2 Flash, MAI Image 2.7, MAI Realtime 1 (from $0.01 per audio minute), the Decision 1.1 routing model and Dragon Copilot Transcribe. Microsoft also announced MAI Licensed Weights, letting customers customise a model and run it in their own Azure environment, and says its Maia 200 chip is in production. All comparisons are Microsoft's own.
industrymicrosoftSource: Microsoft AI: Seven new world-class models at scalable price-performance
US weighs making companies disclose when they use Chinese AI models
Reuters reports, citing three people, that the Treasury Department and the White House Office of Science and Technology Policy have discussed requiring US companies to say publicly on their websites when they use Chinese AI models. Airbnb, which uses Alibaba's Qwen for its customer-service agent, is given as an example. The idea is to discourage reliance on Chinese open-weight models without a ban. It is unclear whether the idea is gaining support or would become a formal rule, and none of the agencies commented. Separately, House committee chairs have written to Airbnb, Anysphere and DoorDash about their use of Chinese models.
policyopen-weightsSource: The Straits Times (Reuters): Trump administration weighs forcing US firms to disclose Chinese AI use
Concept of the day
Over-allotment (greenshoe) option
In an IPO, the company usually gives its underwriters the right, for about 30 days, to buy extra shares at the offering price, typically up to 15% of the deal. Underwriters sell those extra shares at the start. If the price then falls, they buy shares back in the market to support it; if it rises, they exercise the option and buy the shares from the company. Anthropic's option covers up to 147 million shares on top of the 980 million sold, which would add about $12 billion at $83 if fully used.
Models released
| Model | Maker | Size | Notes |
|---|---|---|---|
| Gemini 4 Neon | Not disclosed; 1M-token context | Mid-tier model between Carbon and Hydrogen, with four thinking levels. $3 input, $0.15 cached and $15 output per million tokens. Google reports 61 on the Artificial Analysis Intelligence Index at xhigh and 32% lower cost per task than Carbon at high. A cyber variant is rolling out to vetted defenders. | |
| MAI Code 2 | Microsoft AI | Not disclosed | Everyday enterprise coding model in GitHub Copilot. Microsoft says it reaches 98% of GPT-6.1 Sol's quality at about 40% of the price across four coding benchmarks, and is not tuned for very long agentic work. A cheaper Flash version is also available. |
Local models
Running a 180B mixture-of-experts model on a 32 GB Mac by streaming experts from SSD
Hugging Face's Tarek Ziadé ran Qwen3.8-Flash-Next (48 layers of 512 experts, 10 active per token) on a 32 GB MacBook in Transformers by keeping only the experts the router picks in a 12 GiB device cache and reading the rest from disk, with experts quantised to 3 bits (49 GB on disk). The key trick was a shared slot pool that lets the existing grouped matrix-multiply kernel run on cached experts unchanged. After tuning it decodes about 13.9 tokens a second against 14.8 for the dedicated Strata engine, and prefills about as fast. He still finds it too slow for coding agents, conversion takes about five hours, and the code sits in an unmerged Transformers pull request whose API may change.
localinferenceSource: Hugging Face: 180B Parameters, 32GB of RAM
Try this
Wrap your own tools in the Codex sandbox
OpenAI explained how Codex limits the commands it runs: its default workspace profile lets commands write inside the repo but keeps .git and .codex read-only, allows broad reads, and turns the network off. Enforcement uses Seatbelt on macOS and bubblewrap plus seccomp on Linux. The useful part for anyone running agents is the codex sandbox command, which applies a named permission profile to any command, for example an MCP server started with npx, so it can only write to one folder and has no network. OpenAI's own caveat: broad read access means secrets elsewhere on the machine stay readable, and an agent could still tamper with ignored files such as node_modules that run outside the sandbox later.
securitytry-thisSource: OpenAI: Inside the Codex sandbox
Research
The heads that flip a model's moral answers handle every kind of judgment
Researchers in Luxembourg built 227 annotator-validated minimal pairs that differ in a moral standard, another norm (prudence, etiquette, aesthetics, evidence) or a plain description, and tested 14 open models (Llama, Gemma, Qwen, 1B to 14B). In every model, the attention heads that change a moral answer carried the same share of the effect for non-moral judgments (pooled ratio 1.00), so they look like a general evaluation circuit rather than a moral one. Removing them cut 37-39% of the gap between acceptable and unacceptable ETHICS scenarios, against 1-3% for random heads. The sparse-autoencoder features found mostly tracked the word 'morally' itself. It is a preprint on small open models.
researchinterpretabilitySource: arXiv: Disambiguating Moral Circuits and Moral Features in Language Models
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