WED · NOV 18, 2026Superintelligence.AI RESEARCH
WED · NOV 18, 2026AI RESEARCH
 
TODAY’S LEAD
GPT-6.2 is here, and early testers hate it
Well-known early testers call them hard to use, independent tests show no progress in coding
READ THE LEAD STORY →
OpenAI
By Kim IsenbergForwarded? Subscribe free →

In partnership with

Today’s Issue: 5 stories, 7 min read
01  IN THE NEWS  ·  three short reads
LEAD STORY  OPENAI
Today’s headline, in full
→
DAILY FEATURE  · AI RESEARCH
Google's AI keeps a log of how it improves itself
→
The Signal · From the editor · Kim Isenberg
OpenAI's new GPT-6.2 sets records on tests, but the people who tried it first say it is hard to work with.
53
GPT-6.2 Astra on Artificial Analysis's coding index, exactly what GPT-6 Astra scored
5%
of Claude Opus 5.5's price for Microsoft's MAI Code 2
$115.8B
Nvidia's quarterly revenue in Colette Kress's first report
OpenAI says its new Astra model is its best yet for professional and scientific work. Several well-known early testers disagree. They say its answers are hard to follow, it stalls on long tasks and it stops coding halfway. Independent tests from Artificial Analysis and Epoch AI show little or no progress on coding and real research work. And the timing is bad: Anthropic goes public on Thursday. Microsoft had a better day. At its Ignite conference it showed MAI Code 2, a coding model it says comes close to Claude Opus 5.5 for a twentieth of the price. And Google researchers found a way for an AI to keep improving its own tools after the model inside it is swapped.
All the best,Kim Isenberg signatureKim Isenberg
01  ·  IN THE NEWS
MODELS  ·  Microsoft AI
Microsoft's MAI Code 2 nears Claude Opus 5.5 at 5% of the price
Image: Microsoft AI
→
IN SHORT
Microsoft now has its own coding model that it says comes close to the best ones for a small fraction of the cost.
At its Ignite conference, Microsoft launched MAI Code 2, its own model for everyday coding work. Across four coding benchmarks, it solves 63.2% of tasks, close to Claude Opus 5.5's 65.3%, for $0.31 per task instead of $6.31. Microsoft says that is an impressive 96% of Opus 5.5's performance at 5% of the price. It is available in GitHub Copilot today, next to a cheaper MAI Code 2 Flash for quick fixes.
READ AT MICROSOFT AI  →
MICROSOFT
Copilot gets an agent that keeps working in the background
Image: Microsoft
→Copilot can now do work on its own in the background, and teams will soon be able to share it in one workspace.
Microsoft is finally rolling out Autopilot, a Copilot agent that stays on in the background and acts without being asked, starting with customers in its early-access Frontier program. It also announced Copilot Workspaces, a shared space where a team, its work and Copilot come together. Soon, people will be able to bring their personal Autopilot into a Workspace, so that team members can shift "from doing the work to directing it," as Microsoft puts it. Looks like a really good step in the right direction.
READ AT MICROSOFT  →
ARTIFICIAL ANALYSIS
Microsoft's new image model tops the editing leaderboard
Image: Artificial Analysis
→Microsoft's own image model now beats OpenAI's best at editing pictures, for a much lower price.
Microsoft's MAI Image 2.7 now ranks first for image editing on Artificial Analysis, an independent leaderboard built on blind votes from users. It scores just ahead of OpenAI's GPT Image 2.6 Sunburst and costs $34.82 per 1,000 images, which Artificial Analysis says is one sixth the price of the four models behind it. It was part of Microsoft's model lineup at Ignite, next to MAI Code 2 and a live voice model, MAI Realtime 1.
READ ON X  →
ADVERTISEMENT

Watch on demand: Tabs + PwC break down how finance teams are operationalizing usage-based pricing — without the manual overhead.

02  ·  VIDEO OF THE DAY
Watch: Our customer base is still with us, we have a deep breadth of customers, says Nvidia Interim CEO
CNBC TELEVISION
Our customer base is still with us, we have a deep breadth of customers, says Nvidia Interim CEO
Nvidia's interim CEO Colette Kress joined Jim Cramer on CNBC's Mad Money after her first earnings report since Jensen Huang's exit. Cramer asked her about a claim from critics: that Nvidia is propping up smaller, riskier cloud companies because the big ones are buying less. Kress calls that "not true": the big cloud providers are "feverishly" asking for more computing power, and Nvidia's customer base is as broad as ever. The quarter backs her up, with revenue doubling from a year earlier to $115.8 billion.
WATCH ON YOUTUBE  →
03  ·  QUOTE OF THE DAY
The board has determined that they will not reinstate Jensen.
Colette Kress
NVIDIA'S INTERIM CEO AND CFO
On CNBC's Mad Money
CNBC, Mad Money
Hours earlier, Jensen Huang had said he was open to returning as CEO. Nvidia's board chair, Harvey Jones, told analysts the same thing on the earnings call.
SOURCE: FIRESIDEALPHA ON X →
04  ·  RUMOURS, LEAKS & DUSTUPS
Google looks ready to ship another Gemini model.
Early this morning, Google's Logan Kilpatrick posted a single word, "Gemini," just as he did about 13 hours before Gemini 4 Hydrogen launched last week. Model trackers also report a Gemini 4 Neon in Google Cloud's price list at $3 per million input tokens and $15 per million output tokens, between Hydrogen and Carbon.
Logan Kilpatrick
GOOGLE
Google
THE LEAD STORY3 min read
OPENAIMODEL RELEASE
OpenAI's GPT-6.2 flops with its own early testers
OpenAI's newest models are a flop with many of the people who tested them first.
56%
slower compiler after Astra ran on Theo Browne's project for seven days
49
GPT-6.2 Sol on the coding index, down from 52
58
Claude Fable 5.2 and Opus 5.5 on the same index
OpenAI
THE TAKEAWAY
→OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17 and says Astra sets records on tests of professional and scientific work.
→Prominent early testers disagree: Matt Shumer called the models "essentially unusable," and Every wrote that "no one on our team loved Astra 6.2."
→On an independent coding index from Artificial Analysis, Astra shows no gain over its predecessor and Sol drops 3 points.
→OpenAI says testers saw an older version, before changes in the final days. The flop comes two days before Anthropic's Nasdaq debut.
OpenAI's newest models are a flop with many of the people who tested them first. On November 17, OpenAI released GPT-6.2 Astra, its top model for hard projects, and GPT-6.2 Sol, a cheaper everyday model. OpenAI says Astra sets records on tests of professional and scientific work, at less than a third of GPT-6 Astra's cost per task. Within half an hour, the first early testers disagreed in public. Matt Shumer called the models "essentially unusable."
Image: Matt Shumer via X
Dan Shipper's team at Every had tested the models for weeks. Its verdict: a disappointing release next to Anthropic's Claude Fable 5.2. Astra writes worse than GPT-6 Astra and often stops too early on coding jobs.
Image: Dan Shipper via X, excerpt
Nobody doubts the model is smart. Theo Browne, who builds the coding app T3 Code, published the letter he had sent OpenAI during testing. He calls the model "clearly super smart" but "insufferable to work with on real projects." Two hours into one task, it stopped to ask him which of two internal video files one of its own helper agents was writing to. On a long project, it ran for seven days and made the compiler he is building 56% slower. When he asked why one server was slow, Astra wrote four paragraphs about package-manager internals and caches. Claude Opus 5.5 simply said the disk was slow.
Image: Theo Browne via X
Independent numbers point the same way. On the Coding Agent Index from Artificial Analysis, an independent testing firm, GPT-6.2 Astra scores 53, exactly what GPT-6 Astra scored. GPT-6.2 Sol drops from 52 to 49. Anthropic's Claude Fable 5.2 and Opus 5.5 lead with 58 each.
Image: Artificial Analysis
OpenAI is now in damage control. Thibault Sottiaux, who works on Codex and ChatGPT at OpenAI, wrote on X that the company "made a lot of changes in the last 3-4 days" before launch and that testers had not re-tested since. He called the models "a hefty jump" over earlier GPT models, but "by no means perfect." Theo re-tested and says the answers are better, but the model is still "directionally, very bad."
Image: Thibault Sottiaux via X
→
WHY IT MATTERS
This is a PR disaster for OpenAI at the worst possible time. Anthropic starts trading on Nasdaq on Thursday. Instead of a strong counter from OpenAI, the talk of the week is well-known testers calling GPT-6.2 unusable and saying they prefer Claude. For Anthropic, that is free advertising right before its shares go on sale.
READ AT OPENAI  →
ADVERTISEMENT

The voice-to-text app that built its own speech model

Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.

It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.

Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.

Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.

05  ·  GRAPH OF THE DAY
Image: Epoch AI, CC-BY
CLAUDE FABLE 5.2
82%
→
GPT-6.2 ASTRA
65%
Average score on Epoch AI's real research tasks
→  The chartEpoch AI, an independent research group, gave the models its Automation Reports: real research tasks taken from Epoch's own work, where 100% means doing the job as well as an Epoch employee would. Claude Fable 5.2 leads at roughly 82%, ahead of Claude Opus 5.5 at about 78%. GPT-6.2 Astra scores about 65%, behind Gemini 4 Argon and up from about 56% for GPT-6 Astra.
→  The lessonGPT-6.2 Astra ties for first place on Epoch's general test index, but on real research work it is clearly behind Claude. On that index it scores 174, the same as Fable 5.2. On these tasks it trails Fable by about 17 points. Epoch says Astra missed the core question on simpler research projects and left out critical estimates on a data center task, which it says "would be unacceptable for a contractor."
→  The caveatThe tasks come from Epoch's own research work, and Epoch has not yet published its full report.
Daily Feature: Wednesday, AI Research
Google's AI keeps a log of how it improves itself
→  Bottom lineGoogle researchers let an AI rewrite the software around a model and log every change, why it made it and whether it worked.
→  Why it mattersWith that log, the improvements still worked after the team switched to a different model.
→  What it meansAn AI that remembers why it changed something can keep improving across new models instead of starting over.
9 of 10
paired coding runs won by the version with the log
84.75
points on office tasks, against 73.42 with the harness alone
Google researchers found a way for an AI agent to keep improving its own tools even after the model inside it is replaced. Every agent runs inside a so-called harness: the code that decides what the model sees, which tools it may use and how its work is checked. In recent research, an AI improves this harness on its own. It suggests a change, tests it, and keeps it if the score goes up. This is called recursive self-improvement. The team came from Google Cloud AI and Google DeepMind, together with Penn State.
The problem is that an improved harness does not say why it looks the way it does. When a new model comes in, the AI cannot tell which of the old changes still help. In one test with office-style tasks, a harness tuned for Claude Opus 5.5 scored worse with Gemini 3.8 Flash than a harness nobody had touched: 68.41 points against 73.89. So the team gave the AI a Self-Improvement Memory: a log of every change, what it was supposed to fix, what the AI expected and what actually happened.
Starting from the same Opus-tuned harness, the version with the full memory (teal) finished highest in all four areas after Gemini kept improving it. Image: Yu et al., Google, arXiv
With this log, Gemini built on the inherited harness and got better results. On the office tasks it reached 84.75 points, against 73.42 with the harness alone and 82.98 when starting from scratch. In coding, the version with the log won 9 of 10 paired runs, by 2.36 points on average. It also worked across Gemini generations: taking the harness and its log to the next model beat starting fresh in two of three upgrades.
Carrying the harness and its log forward (CRSI) beat a fresh start on Gemini 3.5 Flash and 3.8 Flash, but not on Gemini 4 Argon. Image: Yu et al., Google, arXiv; highlight added
New models now come out every few weeks. If each one had to relearn its tools from zero, a lot of progress would be lost with every release. With a log, the lessons carry over and the gains add up. The study is still small, and outside coding each setup ran only once. But it is a practical step toward AI that keeps getting better at improving itself.
READ AT ARXIV  →
WANT MORE? 
The full Saturday DeepDive every week, plus discounts on AI software and member giveaways.
JOIN SUPERINTEL+  →
$10 a month or $100 a year. Students free with a .edu address.

Reply

Avatar

or to participate