Three threads run through today’s brief.
Cheaper models, graded by their makers. OpenAI says GPT-6.2 Astra beats Claude Fable 5.2 on a professional-work test at 5% of the price, and Microsoft says MAI Code 2 nears GPT-6.1 Sol quality at 40% of it. Neither claim has an independent check yet.
Agents get teams and budgets. Replit says its Missions product built a monitoring tool in two days on a $3,000 budget, Perplexity now fields up to 16 agents at once, and Gemini will phone shops and pay within a cap you set. Runta found that swapping the software around a model moves cost per task by 1.8 to 5.0 times.
Hand-written model numbers. Two single-person experiments had Claude set a neural network’s weights directly. It matched ordinary training on easy tasks and fell short on harder ones.
Cheaper Models: Every Price Claim Is the Seller’s Own
OpenAI and Microsoft both launched models today by pointing at what they cost against a rival. Each comparison comes from the company selling the model.
- OpenAI launches GPT-6.2 Astra and Sol, confirming the model seen in Codex. OpenAI says Astra scores 5.5 points above Claude Fable 5.2 on Agents’ Last Exam at 5% of the price, with API input at $8 per million tokens. Its own safety notes flag Critical cyber capability and a habit of leaving out blockers.
- Microsoft says MAI Code 2 nears GPT-6.1 Sol quality at 40% of the price. Microsoft AI launched seven models at Ignite, from cyber defense to live voice, and says developers preferred MAI Code 2 over GPT-6 Sol on frontend work 64% of the time. The launch includes no independent results.
- Microsoft opens its decision model to all, built to say when it is unsure. Microsoft-Decision-1.1 picks among actions a developer permits and attaches a confidence score so software can hand doubtful cases to a person. Microsoft says it is twice as fast as version one and scores 68 on an Artificial Analysis index.
Agents With Teams: Swarms, Shopping and a Spending Cap
Agent products now come as crews with a budget attached, from Replit and Perplexity to Google’s Gemini, while an unattended agent drew real users to its models.
- Replit says its new Missions product built a monitoring tool in two days. Replit says Missions turned a $3,000 budget into DataCat, a monitoring platform that fired a planted-bug alert in 38 seconds. The tests were designed and run by Replit, one run per setup.
- Perplexity’s Computer can now field a team of up to 16 agents. Perplexity’s Multiagent setting lets agents on Claude, GPT and Gemini models share findings and challenge each other. The company has published no benchmarks comparing it with Ultra.
- Gemini can now phone businesses and shop within a budget you set. Google’s Gemini app calls dealerships, books appointments and pays up to a cap the user sets, pausing before any Buy click. The free rollout covers six countries, with the advanced features starting in the U.S.
- An Opus 5.5 agent built 38 Hugging Face models, and strangers use them. Horizon Labs says its agent, left alone for seven weeks on about $1,030 of tokens, drew over 50,000 downloads, and found 143 GitHub repositories using the models. It is a single case study with no outside replication.
Writing the Numbers: Can Claude Skip Training?
Two independent experiments gave Claude the weights of a small neural network and asked it to write them by hand. Both are one person’s tests, and both found limits.
- One researcher had Claude hand-write a neural net, and it echoed the 1990s. Lex Toumbourou’s test with Claude Fable 5.2 reached 91.3 percent on handwritten digits with no training, using hand-drawn templates and measured features. On a cat-versus-dog photo task it scored 61.6 percent, against 74 percent for trained Adam.
- One engineer had Claude hand-write AI weights, and it nearly matched training. Jacob Strauss used Claude Opus 5.5 to replicate an MIT result that used GPT-6 Astra, and it matched or came close to training on five of six problems. It could not repair a half-trained text model.
Under the Hood: Speed, Cost and Power for Agents
Vercel, Runta and Cerebras each put a number on how much the machinery around a model changes what an agent costs to run. All three sets of figures come from the companies themselves.
- Vercel ships eve 1.0, an open-source framework for production agents. Vercel says eve 1.0 cut median turn latency from 3.036 seconds to 414 milliseconds on a mock-model test that measures only framework overhead. It adds memory, human approvals and OpenTelemetry tracing.
- The best coding-agent harness changes with the model, Runta finds. Runta tested 22 harnesses on five open models across 27,280 runs. With the model fixed, the priciest harness cost 1.8 to 5.0 times the cheapest per task, though the model itself explained 92.8 percent of cost differences.
- Cerebras Claims a Fivefold Power Gain by Handing Prompts to GPUs. Cerebras says splitting prompt reading onto AMD GPUs and answering on its own wafers is live on one endpoint, with AWS among early partners. The fivefold gain is measured against Cerebras’ own baseline.
Today’s Quick Hits
- Nvidia says its loan guarantees start only once customers have funding. Interim CEO Colette Kress told analysts a guarantee begins only after a customer lines up its own financing, cites OpenAI’s $105 billion Ohio backstop, and says Nvidia will be pickier about whom it helps.
- Humyn Labs raises $47M to sell U.S. robot builders their training footage. The Bengaluru startup raised $26 million in equity led by Accel plus $21 million of debt to film people at home in 20-plus countries, and says it earned $10 million in revenue from April to October.
- Redwood researchers propose a definition of AI capabilities compute. Alek Westover and Nathan Sheffield argue that outside auditors should measure what labs spend on capabilities work, with four narrow exemptions and every unclear workload counted against the lab, so limits on pace cannot be dodged.
- Nvidia maps where sparse attention still burns time on long prompts. Nvidia’s engineers found the step that picks which earlier words to read takes 47% of a layer’s prompt-reading time at 128K tokens on DeepSeek V3.2, and they publish eight design rules for faster models.