Daily Executive Summary

This run closes a two-day incremental window with six new events and three updates. GPT-6.1 Sol is available by API at Sol pricing with near-Astra vendor-reported capability, while Gemini 4 Argon offers a 1M-token output ceiling but remains a controlled preview. The OpenAI Agents API now exposes a resumable hosted-browser contract with origin approval and sign-in handoff. Olmo-core 3 opens a trillion-parameter MoE training stack, and cua-speedrun makes computer-use speed and cost evaluation reproducible. Ollama 0.35.0 was incorrectly treated as prerelease in the prior run and is recovered here as a late discovery. No new trend met the evidence threshold.

Models

  • GPT-6.1 Sol (TRIAL): Available in the API, ChatGPT Work and Codex at $2 input, $0.10 cached input and $10 output per million tokens. OpenAI reports near-Astra coding, computer-use and professional-task capability. Route by completed-task cost and factual errors.
  • Gemini 4 Argon (WATCH): Google raises the output ceiling to 1M tokens and reports 77.9% on DeepSWE v1.1. Access is limited to trusted cyber defenders and post-introductory prices are scheduled to double; do not invest in migration before a public API.

Agent & AI Engineering

  • OpenAI Dots (WATCH): Each Dot has a cloud computer, connects apps and continues work between conversations. Rollout has started for some Pro and Business Premium users, but later pricing and enterprise governance remain unsettled; watch task completion, revocation and audit behavior.
  • OpenAI Agents API Browser (TRIAL): The browser runs in an OpenAI-hosted environment while the application handles origin approval and sign-in fields, with resumable sessions after disconnects. Origin approval explicitly limits which sites the agent may reach; the hard part remains keeping credentials outside model context and gating side effects.
  • Ollama Decision Models (TRIAL, late discovery): /v1/systemone returns choices, probabilities and scores. Calibration asks whether 80% confidence is actually correct about 80% of the time; without it, precise-looking routing thresholds can still be wrong.

Open Source

  • Olmo-core 3 (TRIAL): Ai2 opens a redesigned MoE stack and ran a 1.2T-parameter configuration on 512 B300 GPUs. Expert parallelism spreads experts across GPUs; routing imbalance and communication can erase sparse-compute gains. Published peaks use short random-routing tests, so sustained training needs validation.

Research

  • cua-speedrun (TRIAL): A uniform VM, execution pipeline and agent interface compare computer-use success, speed and cost. Public results show non-monotonic interactions between reasoning effort, environment latency and total time. Code and infrastructure are public for a one-day reproduction.
  • The other 19 new arXiv cs.AI submissions were filtered or deferred for missing artifacts, limited engineering impact or scope mismatch.

Developer Tools

  • Claude Code 2.1.285/286 (ADOPT, updated): Adds provider allowlists and a WebFetch kill switch, and fixes session recovery, Remote Control policy revocation, MCP credential redaction and spend accounting. OS-denied managed settings still warn and continue, so deployments need an external startup check.
  • Gemini CLI 0.62.0 (WATCH, updated): Fixes OAuth refresh-token retention, PTY/ConPTY cleanup and ACP permission-event ordering. Take the stable fixes; the undocumented A2A server remains unsuitable as a production dependency.

Infrastructure

Olmo-core 3 is the clearest infrastructure change. It combines DDP, tensor, pipeline and expert parallelism in an open implementation, but the 1.2T result should not be extrapolated to sustained training without reproduction.

Business & Policy

No business or policy event in the window independently changed model access, chip supply or the open-source landscape. OpenAI's model-distillation campaign disclosure was reviewed as security background, not a new engineering disposition.

Trend Signals

  • Trend #3: Coding agents converge on multi-agent runtimes: GPT-6.1 Sol beta delegation and the hosted-browser Agents API add new production control planes; remains established / High.
  • Trend #7: Agent runtimes separate broad execution from consequence control: OpenAI assigns origin approval and sign-in to the application, while Claude Code disconnects Remote Control after policy revocation; remains strengthening / Medium.
  • The other five active trends were reviewed with no status or confidence change. No new trend met the evidence threshold.

Tech Radar

  • New OpenAI Dots (ai-agent / WATCH): Persistent agents are now a product, but pricing, access and enterprise governance remain unsettled.
  • New GPT-6.1 Sol (foundation-model / TRIAL): Pricing and availability support a real routing trial; verify vendor benchmarks on local traces.
  • New Gemini 4 Argon (foundation-model / WATCH): Strong capability signal, no public API; wait for verifiable access.
  • Agents API Browser (ai-agent / TRIAL): The hosted-browser contract is usable; test credentials, approvals, recovery and audit first.
  • New Ollama Decision Models (ai-engineering / TRIAL): Probability output suits local routing, but thresholds need calibration.
  • New Olmo-core 3 (infrastructure / TRIAL): The open stack is testable; sustained large-scale throughput still needs reproduction.
  • New cua-speedrun (research / TRIAL): The public harness cheaply calibrates speed-and-cost benchmarks.
  • Claude Code (developer-tools / ADOPT): 2.1.285/286 tighten provider, revocation and credential boundaries; rating unchanged.
  • Gemini CLI (developer-tools / WATCH): 0.62.0 stable fixes are useful, but A2A still lacks a contract; rating unchanged.

Current radar: ADOPT 11 / TRIAL 116 / WATCH 109. See the site Insights page for the complete current radar.

Worth Trying

  1. Compare GPT-6.1 Sol, GPT-6 Sol and Astra on 20–50 real traces: success, factual errors, tokens and cost per completed task.
  2. Run one read-only signed-in Agents API browser task with a test account and verify credential boundaries, fail-closed approval, recovery and complete audit.
  3. Rerun 20 tasks with the public cua-speedrun harness; teams with non-B300 hardware can also test sustained Olmo-core 3 throughput at small scale.

Sources