GPT-6.2 Astra/Sol at a third of the cost; Claude Code "Code Mode" (AlphaSignal 2026-11-18)
Why this is in the vault
Daily AlphaSignal record of OpenAI's cheaper GPT-6.2 tiers, an experimental Claude Code mode that batches tool calls via a generated script, and OpenAI's embedded third-party safety assessments.
Mapping against Ray Data Co
Most concrete: Claude Code's experimental Code Mode (Claude writes a small JavaScript program that chains tool calls in parallel, returns a summary, 10-minute script cap, terminal only, enabled via an env var) bears directly on how our cron and sub-agent runs spend tool-call round trips; it is worth a controlled trial in a non-production session before touching settings.json (a gated file). The GPT-6.2 pricing (Astra $8/$40, Sol $2/$10 per million tokens, vendor-reported) extends the price-compression thread behind our model/effort pairing choices, though we run Claude as the daily driver. The OpenAI embedded safety assessments (Apollo noting access still falls short of its bar) are context for the credibility/governance angle, not an action. All figures vendor-reported. Mapping strength: medium; no decision.
Issue contents
Top news (as reported, unverified):
- OpenAI GPT-6.2 Astra and Sol: Astra claimed under one third the per-task cost of GPT-6 Astra, tokens 20% cheaper; Sol near Astra coding at about one fifth the cost; Astra rated High for AI self-improvement; available in the API, ChatGPT Work and Codex.
- Claude Code Code Mode (experimental): JS script chains tool calls, parallel execution, mid-run answers; opt-in via
CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1; not in VS Code or Desktop. - OpenAI embedded safety assessments: Apollo Research, METR, Transluce, Goodfire get deep internal access; Apollo says access is still limited.
- Signals: free uncensored Qwen image model (4.4GB); Datadog survey of 120+ engineering leaders on autonomy blockers (sponsored); Google Cloud multi-agent paper-writing tool; MiHoYo game-playing agent; noisy teacher inputs for distillation; Sarvam India-hosted inference.
- Events promo: two Snyk webinars (Dec 9, Dec 17).
- Deep-fetches: zero (blurbs already specific; no link followed).
⚠️ Sponsorship
Paid placements: Tiger Data (agent memory/traces in Postgres, $1000 credit), TensorWave (457 MW AMD GPUs for OpenAI), Datadog (survey item in Signals). All three are recurring pool members, no new sponsor this issue. Snyk webinars are event promos. Masthead "In Partnership with" slot unresolved. Treat as vendor framing; none touch the editorial items.
Related
- [[2026-11-17-alphasignal-claude-code-iphone-duo-opencode-rust-liquid-d1]]
- [[2026-11-16-alphasignal-claude-code-themes-projects-deepseek-api-cram-energy-model]]
- [[2026-09-23-alphasignal-opus55-gpt6-sol-luna-pricing]]
- [[feedback_delegation_model_effort_pairing]]