Ampersand v0.1 added to BenchGecko · humans& collaborator model at $1.00 and $5.20 per M tokens
humans& released Ampersand v0.1 on 2026-10-20, a collaborator model post-trained from GLM 5.3. BenchGecko lists it at $1.00 input and $5.20 output per M tokens with a 1,048,576 token context.
What launched
humans& released Ampersand v0.1 on 2026-10-20. humans& calls it a collaborator model: it is trained to work with groups of people over long periods, to decide when to speak and when to stay quiet, and to weigh the long term effects of what it says. It is available through the humans& Agents API with no waitlist. The weights are not public.
Ampersand is post-trained from GLM 5.3 with reinforcement learning on long social simulations, where a separate humans& user model called Persimmon plays the people. BenchGecko added the model and the new humans& provider today.
Specs
From the humans& model card and API documentation:
- Base model · GLM 5.3
- Parameters · about 744B total, about 40B active per token
- Architecture · mixture of experts, 78 layers, 256 routed experts plus 1 shared, 8 routed experts active
- Context · 1,048,576 tokens
- Input and output · text
- API name ·
ampersand-v0.1, OpenAI Agents API compatible - Weights · not released
Pricing on BenchGecko
humans& prices Ampersand at $1.00 per M uncached input tokens, $0.20 per M cached input tokens and $5.20 per M output tokens, reasoning tokens included. BenchGecko stores those first-party rates on the model row. The humans& worked example is 1M uncached input, 1M cached input and 1M output for $6.40.
Rate limits depend on the account tier. The Normal tier allows 30 API requests per minute and 16 concurrent executions. The High tier allows 600 requests per minute and 64 concurrent executions. Token prices are the same on both tiers.
No OpenRouter endpoint is live. Compare rates · BenchGecko pricing.
Vendor-reported results
humans& marks these results as tentative and compares them with GLM 5.3. BenchGecko records them as vendor claims. No independent scores or Gecko Tests are attached yet.
- Collaboration Over Time · 12.94 · GLM 5.3 at 9.89
- Collaborative Intelligence teamwork · 59.1% · GLM 5.3 at 54.3%
- CooperBench cooperative success · 39.0% · GLM 5.3 at 17.3%, against 52.5% solo
- CIMemories privacy violation rate · 0.241 · GLM 5.3 at 0.299, lower is better
- SWE-bench Pro v2 Hard-51 · 82.4% · GLM 5.3 at 84.3%
On safety, the model card shows FORTRESS adversarial risk down from 21.4 to 13.8 and multi turn sycophancy down from 77.3% to 62.6% of conversations. WMDP cyber accuracy rises from 80.6% to 89.0%.
Sources
Source: BenchGecko
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.