GPT-6 Sol, GPT-6.1 Sol, Claude Opus 5.5, and Grok 4.7: the late-September 2026 agent routing board

Introduction

The week after GPT-6 Astra’s Critical-cyber launch, vendors compressed the cost of frontier-adjacent coding agents. The useful question is not which model “won” a vendor table. It is which SKU belongs on planner, worker, and verifier lanes, and which still inherits Astra-class cyber controls.

Primary sources: GPT-6 Sol and Luna, GPT-6.1 Sol, GPT-6.1 Sol system-card addendum, Claude Opus 5.5, Grok 4.7, and Copilot weekly notes. Earlier Astra context: GPT-6 Astra Critical cyber routing.

What shipped between September 21 and 29

Grok 4.7 (September 21)

• Larger base than Grok 4.6; longer RL on multi-hour tasks; stronger self-verification; native Grok Bot harness training.
• Standard list $2 input / $6 output per million tokens; Fast variant at 2x speed and 2x price.
• Available in Cursor (desktop, web, iOS, CLI, SDK), Grok Build, and the Grok API.
• Amazon Bedrock availability followed on September 28 with 500K context and four reasoning-effort levels.

GPT-6 Sol and GPT-6 Luna (September 22)

• API IDs `gpt-6-sol` and `gpt-6-luna`.
• Sol list $2 / $10; Luna $0.10 / $0.50. OpenAI states these are 50% below GPT-5.6 promotional rates.
• Work and Codex for Plus through Edu; Luna also on desktop for Free and Go. Not yet in Chat.
• OpenAI later published GPT-6.1 Sol as the newer Sol-class model; keep `gpt-6-sol` pinned until 6.1 Sol is accepted.

Claude Opus 5.5 (September 22)

• First Claude 5.5 family model; ID `claude-opus-5-5`.
• Anthropic: Fable 5.1-level on most work; about 40% lower cost than Opus 5 on typical workloads.
• List $4 / $20; cache reads $0.20 (60% below Opus 5 cache reads). Fast mode $8 / $40.
• Thinking cannot be disabled. Older `computer_20251124` is rejected on Claude API and Google Cloud.
• Safeguards similar to Fable 5.1 for biology and cybersecurity; verification programs for vetted orgs.

GPT-6.1 Sol (September 29)

• API ID `gpt-6.1-sol`. Same $2 / $10 list as GPT-6 Sol, with cached input at $0.10 (5% of uncached).
• OpenAI: near-Astra intelligence for coding, computer use, and professional work at one-fifth Astra standard token prices.
• Preparedness: Critical cybersecurity, High biological and chemical; same safeguards stack as Astra.
• Work and Codex for Plus through Edu; not yet in Chat. No GPT-6.1 Astra launched with it.

How the lanes compare

What this means for developers

• Re-run one eval harness across GPT-6.1 Sol, Opus 5.5 default effort, Grok 4.7, and the current default. Do not mix vendor benches.
• On Claude API, update computer-use tool versions; forced tool use now errors; thinking stays on.
• Pin `gpt-6.1-sol` vs `gpt-6-sol` explicitly. Chat, Work, Codex, and API can roll out on different days.
• In Copilot, enable replacements before October 2 so Opus 4.7 and old Flash IDs do not strand automations.
• Measure cost per merged PR and tokens per resolved task, not only list price.

What this means for businesses

• Planner defaults can move off Astra for many tasks if 6.1 Sol or Opus 5.5 pass quality gates, which changes monthly Copilot and API spend.
• Risk committees should treat GPT-6.1 Sol like Astra for cyber policy even when finance treats it like Sol.
• Bedrock Grok 4.7 lets AWS-centric orgs keep IAM and logging while adding a Cursor-aligned coding model.
• Anthropic’s verification programs matter for life-sciences and cyber-defender teams that need fewer refusals on legitimate dual-use work.

Routing checklist

1. 01
   Inventory model IDs

   List Copilot policies, CI model strings, and SDK defaults that still name GPT-5.6 Sol, Opus 4.7, Gemini 3.5/3.6 Flash, or Kimi K2.7 Code.
2. 02
   Shadow-route GPT-6.1 Sol and Opus 5.5

   Run the production harness at default effort first, then max. Keep Astra or Fable 5.1 as the merge verifier until both pass.
3. 03
   Assign Luna and Grok 4.7 workers

   Put Luna on cheap, high-volume loops. Put Grok 4.7 on long Cursor/Bedrock jobs if CursorBench-style tasks dominate.
4. 04
   Lock Copilot policy

   Enable Opus 5.5, GPT-6 Sol/Luna, GPT-6.1 Sol when available, and Grok 4.7 only for teams with spend owners.
5. 05
   Recalculate unit economics

   Include cache-read rates (Opus 5.5 $0.20, GPT-6.1 Sol $0.10) because coding agents are cache-heavy.

Conclusion

Late September 2026 made frontier-adjacent coding cheaper without retiring frontier governance. GPT-6.1 Sol and Opus 5.5 are the two models most likely to change defaults. Grok 4.7 is the AWS/Cursor price-performance lane. Luna remains the worker. Teams that re-bench once and rewrite Copilot policy will capture the savings. Teams that swap IDs from a changelog will inherit Critical-cyber blast radius at Sol prices.

Sources: OpenAI ; OpenAI ; OpenAI ; OpenAI developers ; Anthropic ; x.ai — grok 4 7 ; github.blog