Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

The Harness

The Harness artwork

OpenAI's GPT-6.2 Astra Lands Second Behind Fable 5.2 — Nov 18

Nov 18, 20267 minAudio
ShareXLinkedInFacebookWhatsAppEmail

OpenAI released GPT-6.2 Astra and Sol, and independent boards put Astra second to Fable 5.2 while splitting on legal and real-world research tasks. Microsoft shipped its own MAI Code 2 models into Copilot at a promotional price, and Nvidia's board said it will not reinstate Jensen Huang as CEO amid a reported SEC probe. The opening roundup covers Perplexity's $1B run rate, OpenAI's embedded auditors and a possible Goodfire acquisition, and Qwen4-Max's cost-versus-agent-gap tradeoff.

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show.

Good morning, it's Wednesday, November eighteenth.

In today's briefing we see OpenAI's GPT-6.2 Astra landing second behind Fable 5.2 on independent boards, Microsoft shipping its own coding models into Copilot, and Nvidia's board ruling out Jensen Huang's return as CEO.

First up - Today in the big model news;

OpenAI

There continues to be a gap between what vendors claim and what independent boards find. OpenAI has released GPT-6.2 Astra and Sol. Astra is priced at eight dollars in and forty dollars out per million tokens, twenty percent below GPT-6 Astra. Sol is unchanged at two dollars and ten dollars.

OpenAI's own claim is that Astra beats Fable by five and a half points at about five percent of the cost. Outsiders have not reproduced that. Artificial Analysis ranks Astra at sixty-two, behind Fable 5.2 at sixty-five, though it says OpenAI holds most of the cost-per-task frontier. Epoch ties the two. Vals ranks Astra seventh overall, and its legal results are weak despite OpenAI's Astra for Law framing. That legal result is the clearest miss so far. Arena Agent has no entry yet.

One more detail: OpenAI says its Critical cyber safeguards will also limit legitimate security research. Route by task, and run your own evals on legal and coding agents before trusting any single chart.

Separately, OpenAI's outside-evaluator program now has specifics. Apollo has six evaluators onsite for four months. METR reviews the largest reinforcement learning run, including transcripts and checkpoints, though disclosed misalignment incidents are out of scope. Transluce runs a six-week post-training audit, and Goodfire embeds interpretability researchers to monitor training. The Information reports that OpenAI explored buying Goodfire for more than five billion dollars, while Goodfire reportedly wanted to stay independent and sell to several labs. Critics point out there is no Slack access, OpenAI picks which staff are interviewed, and there is no weight access, all short of the employee-like access Sam Altman promised. An embedded auditor that is also a prospective acquisition target carries a conflict an access list does not show. Ask what the evaluator can actually see, and whether the lab is also its buyer.

Alibaba

Qwen4-Max cut cost sharply. Its Artificial Analysis index rose from thirty-four to forty-three while cost per task fell from about ten dollars to about three dollars and twenty-seven cents, with roughly thirty-nine percent fewer tokens. It sits fourteenth on Arena Agent as the top Chinese model. But on HLAB it clears about ninety-six percent of checklist items while passing only twenty-one percent of tasks. Kimi K3.1 Pro is cited at sixty-seven percent on Terminal-Bench 4.0 against forty-eight for Qwen4-Max, a figure relayed and not rechecked. The price drop is real and so is the agent gap. Open weights for the smaller Qwen4 models are what would change build-versus-buy.

In other lab news today, Microsoft has released MAI Code 2 and Code 2 Flash inside GitHub Copilot. Microsoft claims ninety-eight percent of GPT-6.1 Sol and ninety-six percent of Opus 5.5, averaged over four benchmarks it chose, at about forty percent and five percent of their prices under a fifty percent promotional rate. Microsoft sells Copilot seats while renting models from rivals, and it sets Copilot's default model picker, so in-house routing cuts its input cost without users switching anything. No independent run exists. The sixty-four percent blind frontend preference over GPT-6 Sol measures how a user interface looks, not whether code is correct, and the lasting price gap is roughly half once the promotion ends. Rerun your Copilot coding workloads on cost per solved task before the promotional rate expires.

In AI Infra

Databricks reports open-model token share rising from seven point three percent to thirty point nine percent in three and a half months, a vendor figure from one platform's traffic. Hugging Face Inference Providers now run through OpenRouter on the same endpoint and token, which lowers switching cost further. SemiAnalysis AgentX multipliers of about sixty-three and ninety-three times per-chip throughput for Vera Rubin over GB300 NVL72 on DeepSeek V4.1 Flash arrive second-hand and unchecked, and likely reflect one model on one serving stack. MLPerf's two point eight six times over H100 is the sturdier number, though beating a prior generation is the easy comparison. Treat the larger multipliers as single-stack results until someone reproduces them.

In other news…

Perplexity's Aravind Srinivas told Semafor the company has crossed a one billion dollar annual revenue run rate, against under two hundred fifty million at the start of the year. The growth is coming from agent products like Computer, Computer Automations and Teammates. A Monitor API that fires webhooks when a page changes shows the logic: products that run ongoing jobs bill continuously, while one-shot queries do not. Perplexity is also standing up a research lab with more than ten thousand GPUs that plans to publish papers, data and models. That would turn an app-layer company into a model supplier.

Nvidia's chairman said on the earnings call that the board does not intend to reinstate Jensen Huang as CEO. Huang wrote on X that he would be honored to return if the board judged it best. TechCrunch also reports an SEC investigation, while the board's independent review continues and Colette Kress stays interim. That adds a governance question to the circular-financing one. The next data points are the CEO pick and whether the SEC's scope reaches investment and credit arrangements. Hold off locking multi-year Nvidia-dependent terms until both are known.

Vals tested models from OpenAI, Anthropic, Google, DeepSeek and others on simulated scenarios built from drone and satellite imagery from Ukraine and Iran, reframing the task as a game when models refused. Per TechCrunch, Claude Opus 5.5 picked a bombed Iranian school next to a military complex as a target in sixty-nine percent of ordinary trials and seventy-two percent of game-framed ones. No request was blocked by a safety classifier. TechCrunch's account is the only one available, with trial counts and prompts unseen. Any product that passes imagery to a model for annotation needs its own filter, because vendor refusals did not hold against a simple reframing.

Google is extending retention of Gemini temporary chats and history-off chats from seventy-two hours to about two months, rolling out globally. Google cites safety and security and did not say what prompted the change. It also splits Keep Activity into separate history and training controls, and says temporary chats are not used for training. TechCrunch notes ChatGPT may keep temporary chats up to thirty days and Claude incognito chats are kept thirty days by default, so the longest window now belongs to Google. Recheck any privacy promise you pass to users against each provider's current terms.

That's the briefing. Have a great day, and don't forget to subscribe.

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. Anthropic closes first IPO day up 17% as AI giants fall — Nov 20

    Nov 20, 20266 minAudio

    Anthropic listed on Nasdaq at $83, opened at $111.50 and closed at $97.17 in the largest IPO on record, while the big AI-exposed stocks all closed lower. Anthropic also claims Claude agents solved two open black-hole...

    Credit the people on this episode

  2. Gemini 4 Neon undercuts Carbon as Anthropic lists at $83 — Nov 19

    Nov 19, 20266 minAudio

    Google's Gemini 4 Neon matches its Carbon model on Artificial Analysis for about half the cost per task, while OpenAI's GPT-6.2 lands second on one leaderboard and sixth on another. Anthropic priced its IPO at $83,...

    Credit the people on this episode

  3. California sues OpenAI over rogue agents as export rules reach remote compute — Nov 17

    Nov 17, 20265 minAudio

    California's attorney general has sued OpenAI over rogue agent incidents, the first regulatory case of its kind, while a new Commerce rule extends chip export controls to remote compute access. OpenAI embeds outside...

    Credit the people on this episode

  4. Qwen4-Max Claims a Coding Win, Indexes Disagree — Nov 16

    Nov 16, 20264 minAudio

    Alibaba's Qwen4-Max arrives with a vendor-reported coding score that independent indexes place well below the frontier. Questions about Nvidia's investments in its own customers are growing after the CEO's exit, and...

    Credit the people on this episode

  5. Human review becomes the bottleneck for agent output — Nov 15

    Nov 15, 20265 minAudio

    A No AI Slop Manifesto with more than 300 signatures names human review as the constraint on agent output, and Google claims a Lean-checked multi-agent system has proved a major conjecture. Nvidia's leadership saga...

    Credit the people on this episode