A 35B Model Matches Trillion-Parameter Rivals by Thinking Longer, Not Scaling Up

01A 35B Model Says It Matches a Trillion Parameters by Running Longer, Not Growing Bigger

Three model releases in the same week share a thesis: the lever for agentic performance is task duration, not parameter count.

The clearest claim comes from Agents-A1, a 35-billion-parameter Mixture-of-Experts model described in a new paper. Its authors say it reaches trillion-parameter-level performance by scaling what they call the agent horizon. They report two axes of scaling: longer trajectories and more heterogeneous agent abilities. To feed that, they built infrastructure linking external knowledge, actions, observations, and verifier outcomes. The resulting training trajectories run to an average length of 45 steps, according to the paper.

That number is the point. A 35B model competing with systems roughly thirty times larger implies the work happened in the chain of actions, not the weights. DeepReinforce's first release, Ornith-1.0, makes the same bet from a different angle, and ships it under an MIT license.

Ornith is built for what its authors call self-scaffolding agentic coding. Developer Simon Willison notes it comes in 9B Dense, 31B Dense, 35B MoE, and 397B MoE variants, built on pretrained Gemma 4 and Qwen 3.5. DeepReinforce claims state-of-the-art results among open-source models of comparable size on coding benchmarks. The open weights matter for cost: a 9B variant can run on hardware a developer already owns, with no per-token bill. The smaller models lean on the agent constructing its own tooling rather than on raw scale.

The frontier labs are narrating the same shift. Anthropic introduced Claude Sonnet 5 as its most agentic Sonnet model, able to plan, use browsers and terminals, and run autonomously. Its own framing is the tell: Sonnet 5 operates at a level that, the company says, "a few months ago required larger and more expensive models." Anthropic says its performance approaches Opus 4.8 at lower prices, launching at introductory rates of $2 per million input tokens and $10 per million output through August 31, 2026.

A mid-tier model closing on a flagship, a 35B open release claiming trillion-class results, and an MIT-licensed model that scaffolds itself all point the same way. Capability is decoupling from parameter size. What determines the ceiling is how long an agent can sustain a task and how well it builds its own scaffolding around the problem.

02Anthropic built Claude a lab bench the same week OpenAI built it an exam

On Tuesday, in front of pharmaceutical executives, biotech founders, and researchers, Anthropic announced Claude Science, a product meant to do for scientific research what Claude Code does for software engineering. According to MIT Technology Review, the company is positioning it as its newest flagship. Given short, high-level instructions, Claude Science can carry out meaningful work on its own, with access to the tools a scientist already uses.

The framing is collaboration. Anthropic says the product sits beside the researcher and runs experiments forward, the same autonomous loop that made Claude Code a daily tool for engineers. It is not pitched as a thing scientists check. It is pitched as a thing scientists work with, from day one, on real projects.

OpenAI spent the same week building the opposite. It introduced GeneBench-Pro, a benchmark that tests how well AI performs in genomics, biology, and scientific research using complex, real-world datasets. The product here is not a collaborator. It is a scorecard. Before a model touches a genome in anger, GeneBench-Pro exists to measure whether it can.

That gap defines the two bets. Anthropic is selling presence in the lab and asking researchers to trust the model by using it. OpenAI is selling a yardstick and asking researchers to distrust the model until it clears the bar. One company wants Claude in the workflow. The other wants AI capability pinned to a number first.

The split matters most to the scientists on the receiving end. A biotech founder who adopts Claude Science treats AI as a lab partner whose output gets folded into ongoing work. A researcher who waits on GeneBench-Pro treats AI as a candidate that has to pass before it earns a role. Same underlying models, two different entry points into the same building.

Neither company has shown the other's hand. Anthropic has not published a benchmark for Claude Science. OpenAI has not shipped a research workbench against which GeneBench-Pro scores would be spent. For now, a scientist deciding how to bring AI into the lab gets two offers: a coworker, or a test.

03The rainbow flowers in those seed listings were never real plants

A buyer browsing eBay, Amazon, or Etsy lands on a photo of a flower too vivid to ignore: petals in impossible color gradients, a bloom no garden has produced. The seeds cost a few dollars. The plant in the picture does not exist. According to 404 Media, scammers are selling seeds for exotic flowers generated entirely by AI, and the three largest consumer marketplaces cannot stop the listings from spreading.

The mechanics are simple, which is why they work. A seller pairs a synthetic image with a generic seed packet, then lists it alongside legitimate horticulture products. The buyer pays for a flower that was rendered, not grown. Whatever sprouts, if anything sprouts, will not match the photo. By the time a germination cycle reveals the gap, the transaction is weeks old and the seller has moved on.

404 Media reports that eBay, Amazon, and Etsy are unable to stem the flood. Each platform runs fraud detection and seller policies, yet the listings keep reappearing faster than moderators remove them. The bottleneck is not detection of a single fake. It is volume. A scammer can produce a thousand distinct flower images in an afternoon and seed a thousand listings before any get flagged.

That economics keeps tilting toward the scammer. Google this week released Nano Banana 2 Lite, which it describes as its fastest and most cost-efficient image model, built for high throughput and scale. The company says it is rolling out across Search, the Gemini app, and developer APIs. The product was not built for fraud. But the trend it represents, cheaper images generated faster, is the same input that makes a marketplace flood possible.

For the buyer, the tell is the image itself. A flower with no botanical name, no nursery selling the live plant, and no second photo of an actual grown specimen is the warning. Reverse image search rarely helps when the picture has never existed anywhere before. The cost lands on hobbyist gardeners and small legitimate seed sellers, whose real listings now compete against synthetic ones that look more striking than anything they can grow.