GPT-6 Astra vs Claude Opus 5 vs Gemini 3.1 Pro: Which Model You Can Actually Deploy This Week

September opened with four labs shipping models in a week, and the press wrote as if every team could swap the new IDs into production on Friday. Most cannot.

GPT-6 Astra began rolling on 3 September for Daybreak cybersecurity partners. Broader ChatGPT and API access followed, but computer-use and the highest reasoning modes are still gated in ways that do not match a typical enterprise contract. Anthropic's Mythos-tier models stay inside Project Glasswing. Google promised a Gemini 3.5 Pro successor for June and spent the summer shipping Flash SKUs instead. Gemini 3.8 Live arrived on 15 September as a voice SKU, not as the missing Pro flagship.

If you are choosing a model this week, you are choosing among the generally available SKUs, not the ones on the keynote slide. That is the whole post.

What you can actually call

Treat that table as a filter, not a leaderboard. Everything in the left column is what gets the quotes. Everything in the right column is what a standard API key can hit on Monday morning.

For the comparison method we use when a SKU is available, see how to evaluate a new LLM in 48 hours. For the July three-way that this post updates, see Opus 5 vs Grok 4.5 vs Muse Spark.

Pricing that actually hits the invoice

Sticker prices as of mid-September, list rates, no committed-use discounts:

Regional data-residency endpoints on OpenAI carry about a 10% surcharge. That is not a rounding error on a support agent doing 2 million turns a month.

Where each available SKU wins

GPT-5.6 Sol, bounded coding and orchestration you already run on OpenAI. If your harness is OpenAI-shaped, Sol is the default this week. The August price cut closed most of the gap with Anthropic on token cost. Computer-use on Astra is the interesting new capability, but it is the wrong reason to freeze a production route waiting for partner access.

Claude Opus 5, long-horizon and instruction-dense work. Still the model we put on unattended overnight jobs and on extraction that has to follow a rule rather than invent a plausible field. Fable 5.1 is the cost lever when the same prompt is reused: cache reads dropped 75% on 1 September, which is a real invoice change for RAG-shaped and agent-loop traffic. Details in Claude Fable 5.1 cache pricing.

Gemini 3.1 Pro / 3.8 Flash, multimodal and Google-stack shops. If the rest of the product already lives in Vertex or Gemini Enterprise, switching labs for a 3-point SWE-bench delta is vanity. 3.8 Flash is the SKU to shadow if your agents are tool-heavy and latency-sensitive. 3.8 Live is a voice product; do not mix it into this table. See Gemini 3.8 Live for voice agents.

Failure modes the launch posts skip

Gated SKUs silently no-op. A wrapper that retries Astra on 403 and falls through to Sol is fine. A wrapper that treats 403 as "the model is down" pages the on-call at 2am. We have seen both.

Thinking-on-by-default truncates. Opus 5 still burns max_tokens on hidden reasoning unless you set it. Swap the model ID without revisiting the parameter and you ship empty completions that look like provider outages.

Export and residency. US Commerce controls from June still restrict some Anthropic frontier SKUs for foreign-national staff, including teams in India. If that is your bench, read export controls and Claude Fable 5 before you pick a default.

Flash intro pricing expires. Google has already said 3.8 Flash list rates double on 1 January 2027. A build that only closes because of intro tokens will miss its 2027 budget.

The routing table we would run this week

No overall winner. Anyone publishing one is averaging away the distribution you actually have.

How to decide without waiting for Mythos

Keep the harness provider-agnostic. OpenAI-compatible plus Anthropic is enough; Muse Spark and Grok already sit on those shapes.

Run 48 hours of shadow traffic on the available SKU, not the gated one. Measure cost per completed task, retry rate, p95 latency, and the failure mode you can name in a sentence.

Revisit the table when Astra computer-use is on your contract, when Mythos is generally available, and when Gemini 3.5 Pro actually ships. Until then, ship on what you can call.

Frequently Asked Questions

Should we wait for GPT-6 Astra computer-use before starting an agent project? No. Build the harness, evals, and tool contracts on GPT-5.6 Sol or Opus 5. Drop Astra onto a shadow route when your account can call it. Waiting for a gated SKU is how pilots miss two quarters.

Is Claude Fable 5.1 a replacement for Opus 5? On cache-heavy, well-specified work, often yes. On overnight autonomy and literal instruction-following, we still pay Opus 5. Shadow both; do not take our word for the split on your traffic.

Does Gemini 3.8 Flash replace 3.1 Pro? For agentic coding and tool loops, it is the SKU to try. For long-context multimodal documents, keep 3.1 Pro in the matrix until you have numbers. Intro Flash pricing is not the 2027 price.

How do we compare models we cannot access yet? You do not. Put them on a watch list. Compare the SKUs in the "route today" column. Benchmark tourism on models your key cannot hit is a waste of an engineer's week.

---

FoundrySoft designs model routing, eval harnesses, and production agent stacks for teams that have to ship this quarter. See AI agent development or start a conversation.