In the space of three days the two labs whose models we use most both shipped their flagship. Anthropic released Claude Fable 5.1 on 1 September. OpenAI released GPT-6 Astra on 3 September. Both cost $10 per million input tokens and $50 per million output. Both claim to be the best software engineering model available. Both became the default in their maker's coding agent within the week.
That leaves anyone running a build with a question that benchmarks cannot answer on their own: which one, for what? Here is how we are answering it.
What OpenAI Shipped
Astra is OpenAI's most capable model, and the company is pitching it at computer and browser use as much as at code. Greg Brockman called it "our most intelligent and, also very importantly, our most aligned model yet", and went further on stage than OpenAI usually does about how close this is to general intelligence. Access went first to a cybersecurity program, with paid plans and the API following.
The technical detail that has the research community arguing is a technique OpenAI calls opaque recurrence. The model carries reasoning state between steps that is not exposed as text, which is the part that makes it efficient and the part that makes it harder to audit. Jakub Pachocki's framing was that more capable models can do harder tasks with fewer language tokens. Critics point out that a chain of thought you cannot read is a chain of thought you cannot check.
The numbers, as reported: a perfect score on ExploitBench against Sol's 78.5 percent, 99.2 percent on a binary reverse engineering benchmark, and 100 percent on an eight-needle retrieval test up to 512,000 tokens of context. On Artificial Analysis's independent intelligence index it scores 61 against Fable 5.1's 66, while on the same site's coding agent index it lands two points above Sol at similar cost and reaches Fable 5's score at less than half the cost. In other words, the benchmarks disagree, and which one you quote depends on what you want to be true.
Launch day was also the day the Claude API, claude.ai and Claude Code went down for about three hours, alongside outages at ChatGPT and Grok. Everyone had a bad Thursday.
What Anthropic Shipped Two Days Earlier
Fable 5.1 kept its price and changed the part of the price that matters for agents. Cache reads dropped from $1 to 25 cents per million tokens, a 75 percent cut. Anthropic's estimate is that typical workloads get about 25 percent cheaper and highly agentic ones about 45 percent, because a long Claude Code session re-reads the same context on almost every turn. The model itself moved from 42 to 55.8 percent on Terminal-Bench 4.0, and its cyber classifier fires about 60 percent less often on ordinary work, which anyone who spent July arguing with it about a debugging session will appreciate. It became the Claude Code default the same day.
What Actually Decides It
We have had both on the same client repository since Friday, with the eval set we build from every Flight Plan's acceptance criteria. Here is the list we work through, in the order it matters to a build.
The harness you already run. Astra is the default in Codex, Fable 5.1 in Claude Code. Each model is tuned for its own agent, and Astra's efficiency depends on a harness that preserves its hidden reasoning state between requests. OpenAI's own adapter does that. Running Astra through a generic harness gave dramatically lower results on at least one benchmark. So the first question is not "which model" but "which agent", and for a client build that is decided by the repository's rules, skills and hooks, which we keep tool-agnostic precisely so we can switch.
Cache economics. Sticker price is identical. Effective price is not. A Claude Code session on a large codebase is cache reads by volume, and Fable 5.1 just made those four times cheaper. Astra's efficiency shows up as fewer output tokens per task. Which wins depends on whether your work is long sessions over a stable codebase or short bursts of hard reasoning. Our expectation going into the eval is that the cost per completed ticket lands closer than either sticker suggests.
Availability and region. Fable 5.1 is on Bedrock, Vertex and Microsoft's cloud from day one; for an Australian client on Bedrock it currently routes through global cross-region inference rather than Sydney, which is a data residency note in the architecture overview, not a blocker. OpenAI lists AWS among Astra's launch channels. And both vendors had an outage this week. The fallback chain we wrote about in July is not a hypothetical.
Auditability. Opaque recurrence is a genuine trade. For an assistant that summarises a document, nobody cares. For a workflow that decides whether to issue a refund, a client's compliance function may want to see the reasoning, and a model that keeps it hidden is a harder conversation. We flag it per capability rather than per vendor.
The eval on your work. Everything above is context. The decision is the eval: last month's tickets, both models, same harness rules, scored on the acceptance criteria and the cost. Our expectation is that Fable 5.1 wins on multi-file changes that need the whole codebase in view, and Astra wins on isolated tasks with hard logic and a tight budget. The eval is there to prove that wrong, and it will not split the same way for the next client, which is why it runs every time.
How This Changes a Flight Plan
Not much, and that is the point. Since July every architecture overview names a capability, a quality bar, a primary model and a fallback chain. This week the top of that chain has two candidates at the same price, and the eval set from the Flight Plan is what picks between them. A founder does not need to hold an opinion on opaque recurrence. They need a crew whose process produces the answer for their product and can produce a different one in six months.
What we would push back on is the framing in the launch coverage that one of these models supercharges everything. The model that helps most is the one that is cheapest for the work you actually do, inside the agent you actually run, on the day it is actually available. Two of those three change monthly.
Which Frontier Model Should Your Product Use?
At Dev Force One, the answer comes from an eval on your own work, not from a launch post, and the process is built to change its mind when the next release lands.
What you get:
- A Flight Plan eval set built from your acceptance criteria and run against every candidate model
- A tool-agnostic build: rules, skills and hooks that move between Claude Code and Codex without a rewrite
- A fallback chain with real costs, including cache economics and regional availability
Choosing a model for a product, or unsure the one you have is the right one? Book a Pre-flight call and we will run the numbers together.