OpenAI posts 722 unverified math manuscripts, and Haiku 5.5 cuts small-model prices
OpenAI dumps 722 unreviewed math manuscripts on the world, Haiku 5.5 drops small-model prices by roughly 90%, and a few honest lessons about what makes agents and datastores reliable.
Most of today’s stories come down to checking. Who checks the proofs, who checks the click, who checks what an agent claims about itself. The answers are less glamorous than the demos.
1 · Top story
OpenAI publishes 722 math manuscripts and says they’re still being verified
OpenAI released 722 manuscripts, grouped into 372 families, produced by an unreleased internal model working through roughly 4,000 research problems at about three hours of ChatGPT Pro compute each. The repo says the results sit at “different stages of verification”: many have Lean formalizations, some don’t, and the authors say unformalized ones “could have issues.” They frame it as an evaluation of open problems, not peer-reviewed work.
Headlines about solving famous conjectures are coming from commentators, not from that repo, so I’d wait for the formal checks. The part for builders is the pattern: output cheap to generate in bulk, expensive to verify, with a formal checker as the only thing that scales. If your agents produce code or analysis at volume, your verifier is the real product.
2 · Models & features
Haiku 5.5 cuts small-model prices to $0.10 in, $0.50 out
Anthropic’s new small model is priced at $0.10 and $0.50 per million tokens up to 100k prompt tokens, rising to $0.50 and $2.50 beyond that, against $1 and $5 for Haiku 4.5. Cache reads drop to $0.01. It’s the first Haiku with an effort setting (low through max), and Sonnet 5.5 cache reads are also halved. Anthropic reports 72.4% on an offline OSWorld subset and 39.2% on Terminal-Bench 4.0, against 70.6% for Sonnet 5.5 on the latter.
That Terminal-Bench gap is the useful number: Anthropic itself points Haiku at compaction, summaries, classification and subagent work, not hard agentic coding. Two things to check before you re-cost a pipeline: the tokenizer uses somewhat more tokens per task than Haiku 4.5, and the 100k step matters if your prompts run long. The benchmarks and customer figures are all Anthropic’s or self-reported.
3 · Agents
Stacklok says agent harnesses should live in the cloud
Craig McLuckie and Joe Beda of Stacklok argue that harnesses built for desktops break when lifted into containers: lifecycle is awkward and pausing an agent for human input is worse. Their open-source harness keeps the agent loop independent of client, model provider, state store and execution environment, with tool calls and bash split off from the loop and session state moved out of local files.
It’s an interview with a vendor selling the control plane, so read it as a design argument, not evidence. Still, the separation is the thing to steal: if the loop, the sandbox and the state can’t be restarted independently, you can’t audit or budget them either.
4 · Research
A web agent’s biggest bug was a coordinate offset
The WebFovea paper treats each agent step as a round trip: parse the reply, execute the action, report feedback, give the model an observation. A bug placing clicks at about three-quarters of the intended position was costing most of the score. Fixing it moved 24 dev tasks from 8 to 20 successes, and the share hitting the step limit fell from 67% to 4%. Same model throughout, 2nd place at 57.0/100 in the WebRetriever Challenge 2026.
Small samples and one model, so treat it as direction, not proof. But when your eval says the model failed, check the harness first. Also worth noting: feedback notes alone didn’t stop idle loops. Executor-enforced budgets did.
5 · People worth following
Martin Kleppmann launches a centre for cryptographic trust between agents
Kleppmann announced the Centre for Cryptographic Trust Infrastructure, funded by a £2.4m ARIA grant with Cambridge, Ink & Switch and others. The bet is that zero-knowledge proofs can let one agent prove to another that it’s telling the truth, using signed records like invoices or sensor data.
There’s no website, API or tooling yet, so this is a research direction, not something to plan around. It’s worth watching because most agent-to-agent designs today lean on identity and policy, and nobody has a good answer for “prove it” without handing over the data.
6 · Architecture
Grab halved p99 by changing its data model, not just its database
Grab moved its anti-fraud Counter Service to Aerospike and reports about 50% lower p99 reads, data down from roughly 3 TB to 1 TB, and 45–50% lower cost per node. The catch was a primary index of about 64 bytes per record, so it redesigned from one row per time bucket to one record per counter holding a timestamp-ordered map, moving time filtering to the client. It ran shadow reads at 5%, 20%, 50% and 100% before cutting over.
The numbers are Grab’s own, for one workload. The transferable lesson is that the index cost, not row size, shaped the schema, so model that before migrating any key-value store.
Nothing here needs doing before lunch. Go and check one of your own harnesses instead.
Researched and drafted with AI, checked against the linked sources. How we use AI →