Is a cheaper coding model safe in production? SWE-2's 64% cost claim

You are deciding whether to route production coding traffic from Fable 5.1 to SWE-2. The trade-off is a small quality spread on Cognition's FrontierCode 1.1 for a large price spread: SWE-2 scored 50.0%; Fable 5.1 reached 50.9%, and GPT-6 Astra reached 53.3%. The arithmetic is simple: SWE-2 is 64% cheaper than Fable 5.1 on Cognition's FrontierCode 1.1 Main comparison, so a 100-unit current spend drops to 36 units. If the quality spread is on a task you do not run, the price spread is real. If it is on a task you do run, the price spread is a discount on risk. The catch is that the invoice, the harness, and the task mix all move the number.

The spread is the entry fee

Cognition made SWE-2 available on September 10, 2026. The SWE-2 vs Fable 5.1 comparison is the test case: the quality row is close, and the price row is not.

A cheaper near-frontier model is safe only when its cost advantage survives your serving path and no production-critical benchmark shows a disqualifying gap. The substitution rule is a sequence, and the order matters. A clean cost line cannot rescue a row you cannot audit, and a strong benchmark cannot rescue a model you cannot reach before you route traffic.

The checks decide whether the saving is real

• Verify benchmark provenance. SWE-2's rows depended on Cognition's execution setup, using public results when present and otherwise Devin CLI runs at the best reasoning-effort setting. The published rows included 92.8% on Terminal-Bench 2.1 and 73.0% on DeepSWE, with the DeepSWE row below the best verified Muse Spark 1.3 score of 75.4%. The audit table names the harness, the effort setting, and whether each row is public or provider-run.
• Rebuild the cost basis. Cognition reported that SWE-2 medium used 58% fewer turns and cost 81% less than SWE-1.7 on FrontierCode 1.1 Main. The Efficiency Frontier paper's HotpotQA evaluation showed that a deployment-aware strategy cut effective token use by about 25% while keeping F1 near 0.78, and that amortized memory compression cut token cost by more than 50% versus full-context prompting in higher-performance settings. Do not treat a provider headline as your invoice. The cost estimate includes turns, context, retries, and your serving stack.
• Confirm availability. SWE-2 was available in Devin Desktop, CLI, Web, and Fusion at launch. It was free for Pro, Max, and Teams subscribers for one month. Cognition had not yet released a separate API, pricing details, or model weights. A free month is a trial window, not a production SLA. The supported path runs from your router to the model, with a named owner and a fallback.
• Find disqualifying gaps. SWE-2 scored 27.30% on Terminal-Bench 4.0, while the best verified score shown for that benchmark was Claude Mythos 5.1 at 60.90%, a 33.6-point gap. The FrontierCode row is not the Terminal-Bench 4.0 row. If your production traffic includes long-horizon terminal work, that gap can be the whole decision. The task list gives each routing bucket a pass threshold and a measured score.

The routing rule is a threshold, not a preference

Route only when every check passes together, and keep the decision tied to the workload, not the model name. The benchmark row traces to a harness you can reproduce or at least audit. The cost saving survives your serving path, not just the provider's comparison. The model is reachable through a supported interface, not a free month or a desktop app. The gap on your critical task stays inside the threshold you set before the test.

If one check fails, keep the traffic on the current model, write down the failed check, and re-test after the gap or cost basis changes.