The Pulse, Thursday 9 October 2026
OpenAI withdrew three math papers after community pushback, while Anthropic cut Sonnet 5.5 cache reads in half and announced API credits for Max/Team plans. Claude Code is being used to discover planets, and a 16.9 MB speech model just showed up on Hacker News.
Models
StepFun Step 5 Preview
A 27B active / 600B sparse MoE with 1M context, optimized for agentic work and software engineering, now on OpenRouter.
Large context and agentic focus fits our Hermes workflows; if it runs well on agent tasks it's worth testing against Sonnet and Sol for code and analysis work.
OpenAI GPT-6.1 Sol Ultrafast
Near-Astra intelligence at up to 8x faster speeds than Sol Standard, rolling out to the API, Codex, and ChatGPT Work.
8x faster Sol means our Codex and Claude Code jobs finish sooner, cutting latency on reviews, code generation, and agent work.
Research
From Retrieval to Reconstruction: Constructing Evolvable Cognitive Memory for Long-Term Dialogue
A memory system that distinguishes source-attributed beliefs from unattributed facts, enabling long-term dialogue agents to reason reliably over extended interactions.
EverRaise voice agents and our own Mnemosyne work could benefit from separating what the agent learned from what the user stated, especially for multi-session calls.
DuplexAgent-RSI: Recursive Harness Improvement for Full-Duplex Voice Agent Collaboration
A full-duplex model stays on the live channel while complex reasoning and tool use are delegated asynchronously, solving the latency-vs-capability trade-off in voice agents.
EverRaise runs Vapi voice agents, and this pattern—lightweight duplex front-end plus async delegation—maps directly to what we build for call grading and lead nurturing.
All Verdicts are Not Equal: Rethinking LLM Judge Reliability
A comprehensive reliability audit of LLM-as-a-Judge across six models, four benchmarks, and varying prompts, showing systemic variance often treated as ground truth.
We use LLM judges in the wehshi review panel and for evaluating prompt outputs; understanding their brittleness helps us calibrate confidence in automated verdicts.
SWE-Journey: Towards More Realistic Evaluation of Coding Assistants through Long-Horizon, Multi-Turn Interaction
A benchmark for coding assistants that spans long chains of development work in evolving repositories, with repeated requirement clarification—closer to real-world Codex and Claude Code use.
Our team uses Claude Code and Codex daily; a benchmark that mirrors actual multi-turn, evolving-repo workflows is a better calibration tool than one-shot tasks.
Releases
LangChain 1.4.4
Fixes retry logic for summarization middleware on context overflow, plus dependency bumps.
We use LangChain for retrieval and agent flows in Pipeline's clients; retry on overflow is a common edge case in longer documents.
LangChain Core 1.6.9
Adds support for passing a callable to with_retry.
Custom retry logic is useful for rate-limited APIs; if we're hitting OpenRouter or vendor endpoints with dynamic backoff, this gives us the hook.
Langfuse 4.55.0
Adds support for OpenAI decision models in evals, FIPS-mode base images, skill imports from ZIPs, and batch score processing.
We trace with Langfuse; batch score processing and decision model support fit our review and agent eval workflows.
Hacker News
OpenAI Withdraws 3 Math Papers
OpenAI retracted three mathematical results from its public repo after community scrutiny.
When we rely on vendor benchmarks or published research for model selection, this is a reminder to verify claims ourselves before betting production work on them.
Whistle: Speech to Text in 16.9 MB
A 16.9 MB speech-to-text model, tiny enough to run anywhere, with competitive accuracy.
EverRaise voice agents process transcripts in real time; a model this small could run on-device or in edge functions with near-zero latency.
I think I found a planet nobody knew existed. I used Claude Code to find it
A user claims Claude Code helped discover a previously unknown planet through data analysis.
This is one data point on how far coding agents have come—enabling domain work (astronomy data crunching) for non-domain experts.
OpenAI annualised revenues $20B less than previously signalled
OpenAI's annualized revenue is $20 billion below earlier projections.
No direct use for us yet—but it's the talk of the day and signals the AI market's volatility matters for pricing and vendor stability.
People on X
Anthropic adds monthly API credits to Max and Team plans
$100 on Max 5x, $200 on Max 20x, and up to $500 pooled on Team, usable on any Claude model.
Our team runs on Claude subscriptions; API credits let us script more work without separate billing, useful for cron jobs and Hermes automations.
Anthropic cuts Sonnet 5.5 cache reads in half
Cache reads now $0.10 per million tokens (was $0.20), making Sonnet 5.5 ~20% cheaper on agentic work.
We use Sonnet 5.5 for agent work and PR reviews; 20% cheaper cache reads means our wehshi panel runs cost less per review.
Hermes Agent now in the Microsoft Store
One-click install for Hermes Agent on Windows, with more updates for the Windows ecosystem coming.
We run Hermes Agent as Botterfly; easier distribution helps onboard clients and partners who want to try the agent on their own machines.
Building effective agent automations with Claude Managed Agents
A guide to deploying agents that read Slack/GitHub on a schedule with credentials and memory, using Claude Code and the new API credits.
This is the same pattern we use for The Pulse and other cron jobs; the official guide might have shortcuts we missed.
Releases: Next.js shipped only canaries in the window, so the three LangChain and Langfuse releases above are the notable ones.