Issue 9: Gemini 4 Argon and Meta's Muse Wonder arrive
Two frontier-class models arrived in one week: Gemini 4 Argon is now generally available, and Meta’s Muse Wonder family scores near the top of Artificial Analysis’s index. The second thread is the infrastructure that agent-scale work runs on. Modal passed $1B ARR. Anthropic replaced git with Scion, its own version control, after daily merges grew more than tenfold. Claude agents can now export OpenTelemetry traces, in beta. On the OpenAI side, GPT-6.1 Sol gets a 500k default context in Codex and Work, and GPT-5.5 has left Codex for users signed in with ChatGPT.
- Gemini 4 Argon goes GALive on the paid Gemini API, Google AI Pro and Ultra. Artificial Analysis puts it level with GPT-6 Astra at about 60% of Astra's cost per task at introductory prices: $2/$10 per million input/output tokens, $4/$20 after 30-11-2026. AINews says most benchmarks confirm Google's claims.
- Meta's MSL ships Muse WonderWonder ships to the Meta Model API and Muse Code, Wonder Full to Muse Code subscribers only. Artificial Analysis scores them 52 and 56 on its Intelligence Index v4.3 at max reasoning effort, up from 48 for Muse Spark 1.3; Opus 5.5 leads at 58.
- Modal passes $1B ARR with a $750M Series DAccel led the round; per AINews, the $15.75B valuation tripled in five months. Modal says usage tripled since its Series C and the company grew over 15x year over year; Sandboxes, launched in 2023 for running LLM-generated code, alone has a $300M run rate.
- Tell the AI its budgetMatt Shumer's one-minute post: agents quit early with half-baked work or triple-check past budget. His one-sentence prompt addition states the token budget and asks the agent to track spend and finish within it; he says it works in Claude Code, Codex and Cowork, not chat.
- CLI --help as LLM contextGeoffrey Huntley tiers developer-tool companies by whether models know their tool from training: S-tier means a model uses the CLI without skills or an MCP server. Newer companies should build a strong CLI, tune --help for fewest tool calls and publish docs agents can search.
- Nobody should type code by handHuntley distills four loop-engineering episodes with Josh Wulf: the operator's judgment is the limit, not the models; delete skill packs, as unverified ones may degrade the model; test the whole system instead of relying on human code review; engineers keep vision, verification and agency.
- Agent observability on Claude PlatformOpenTelemetry trace export, in public beta, covers each agent turn, model call, tool use and MCP server call with latency, tokens and IDs. The SDKs export automatically; Managed Agents and the Messages API need a Claude Console destination: Datadog, Honeycomb, Langfuse or any OTLP endpoint.
- Anthropic rebuilt source controlDaily merges rose from about 2,500 in May to over 30,000 by September, outgrowing git's offline, full-copy design. Seven engineers built Scion for always-online clients, nearly all agents and jobs, and moved the monorepo off git in five months; they rate Claude the better coder.
- GPT-5.5 retires from CodexAt noon PDT (19:00 UTC) on 14-10-2026 GPT-5.5 left Codex under ChatGPT sign-in; the OpenAI API and API-key Codex sessions keep it. Update workspace defaults, saved model settings, managed configs, custom agents, scheduled tasks and scripts still on gpt-5.5; OpenAI recommends GPT-6.1 Sol where offered.
- GPT-6.1 Sol: 500k default context in Codex and WorkThrough Codex and Work, the default context rose from 272k to 500k tokens, per OpenAI's Tibo, on day 7 of a 28-day series. A forum moderator's summary adds: long-context pricing starts past the new threshold, and default 90% compaction keeps sessions under it unless changed.