OpenAI tells developers to run GPT-6 Astra at low reasoning, and one developer's test cost $25.67 against Sol-high's $31.79

The New Stack reports that OpenAI's Codex engineering lead Thibault Sottiaux told developers on X over the weekend to run the new GPT-6 Astra at low reasoning, saying it beats GPT-5.6 Sol on high. Astra tokens cost 2.5 times what Sol tokens cost, so the advice matters: the piece pairs Sottiaux's quote with two independent tests that pull the cost picture in opposite directions.

Astra is priced at $10 per million input tokens and $50 per million output tokens against $4 and $20 for Sol. Artificial Analysis currently scores Astra-low at 49 on its Intelligence Index and Sol-high at 48. Astra-low returned its first token in 2.53 seconds against 11.87 seconds for Sol-high, per the same source.

Shinpr's test: fewer requests, lower bill

Developer Shinpr ran the same codebase through three phases, analysis, implementation and review, at Sol-high and at several Astra reasoning levels. The New Stack reports the numbers: Astra-medium handled implementation in 80 requests, less than a third of the 238 Sol-high needed, and processed 11.1 million input tokens against 37.8 million. The full three-phase Astra-medium run took about 51 minutes and cost an estimated $25.67. Sol-high took roughly 75 minutes and cost $31.79.

Turning Astra up to high made it worse. That run took 77 minutes, cost $37.23, and Shinpr says the review missed a startup bug that the medium run caught.

ARC Prize: more reasoning, lower total

ARC Prize saw the opposite shape on ARC-AGI-3. Astra scored 17.5 percent at low reasoning, 38.6 percent at medium, 54.8 percent at high and 62.7 percent at max. ARC Prize spent $38,166 at low, $48,090 at medium, $40,705 at high, and $26,098 at max: the cheapest run was the one at max. Higher reasoning per decision meant fewer wrong turns, and fewer wrong turns meant fewer tool calls to redo.

OpenAI's own numbers from Astra's launch fit the same pattern. On Terminal-Bench 4.0 the company reports Astra at 57.9 percent against Sol's 37.3 percent while costing about 9 percent less per task; on GPQA Diamond, 94.9 percent against 94.6 percent at an estimated 37 percent lower cost. Those are vendor benchmarks and should be read as such.

A dial you can move mid-conversation

Astra adds a configuration_update mechanism that lets an application change the reasoning setting between responses without editing the original request-level configuration. The New Stack, citing OpenAI's docs, notes it only works with Astra in standard, single-agent requests today.

That mechanism is the reason Shinpr and ARC Prize can both be right on the same model. Routine work stays at low reasoning, then a failed test, an unexpected tool response or a hard debugging step promotes the next turn to high, and once resolved the agent drops back down. Which setting wins on cost depends on how often the model would take a wrong turn without the extra thinking.

The takeaway for a team weighing the upgrade is to test both models across reasoning levels on the workload that matters. Run the same task set at low, medium and high, and count both the requests and the total tokens for the whole task rather than the per-token rate. On short work like Shinpr's, medium can beat high on cost and quality at once. On long agentic runs like ARC Prize's, one more thought each step can save several actions later.

Source

OpenAI's new model costs 2.5x more per token, and developers are saving money anyway, The New Stack, 7 September 2026. Primary developer reports referenced: Shinpr on Dev.to and ARC Prize on GPT-6 Astra.