GPT-6.1 Sol Arrives With Near-Astra Performance at a Fifth of the Price

OpenAI announced GPT-6.1 Sol at DevDay on September 29, and the positioning is the whole announcement: near-GPT-6 Astra performance at roughly a fifth of Astra’s price. The spec sheet backs the framing: $2 per million input tokens, $0.10 for cached input, and $10 output, against Astra’s $10/$50, with a context window around 1.05 million tokens and 128,000 max output tokens. A new Ultrafast tier runs at 300 tokens per second for latency-sensitive agent loops. What makes this release unusual is not the spec, it is the date: GPT-6 Sol launched on September 22 at this exact price. Sol 6.1 is a capability refresh seven days into a generation’s life.

The Weekly Point Release Is the Real Signal

Model generations used to be annual, then quarterly; GPT-6’s flagship tier has now shipped twice in eight days. The economics of the current price war explain why: the frontier repriced three times in seventy-two hours last week, Anthropic answered with Opus 5.5 taking most benchmark crowns and Sonnet 5.5 matching Opus-class workloads at $2/$10, and Grok 4.7 started the whole cascade at $2/$6. In that market, a capability gap between your $2 tier and the competitor’s $2 tier is a churn risk measured in days, and a training-run improvement you would once have baked into GPT-7 instead ships as 6.1. The practical consequence for builders is that model choice is now a weekly procurement decision, not an architectural one, and any pipeline hard-coded to a model name needs a router.

What the Claim Does and Does Not Say

The “near-Astra at a fifth of the price” framing is deliberately careful. Near is not equal, and Astra retains the places where its extra reinforcement budget shows: long autonomous runs, where AutomationBench favored it against even Opus 5.5, and the restricted cyber capability tier that no cheaper model ships with. Independent benchmarks for 6.1 specifically are not yet available; the launch coverage (VentureBeat, Vellum, DataCamp) leads with the price-performance claim and OpenAI’s own numbers, and the few third-party figures circulating for GPT-6 Sol show how inconsistent cross-source scoring is for this generation. The defensible reading: 6.1 narrows the gap to Astra on the workloads most buyers actually run (coding, knowledge work, tool use at high effort), keeps the price floor, and leaves the frontier tier to Astra until the next Astra.

The 1M Context and Ultrafast Tier Are Aimed at Dots

Two specs reveal who this model is for. The 1.05 million token context window matches the always-on agent requirement: a Dot running for days accumulates project state that 400k-window models churn out of. And the Ultrafast tier at 300 tokens per second targets interactive loops where perceived speed is the product. Dots, announced at the same event, run on Astra today, but the spec-sheet geometry suggests the plan: Astra anchors the premium, 6.1 Sol becomes the workhorse for always-on agent fleets where a fifth of the price at near-Astra quality turns always-on from a Pro-tier luxury into a mass-market product. This is the Sonnet 5.5 playbook executed in reverse: Anthropic proved the mid-tier can match the flagship; OpenAI is now proving the flagship’s quality can descend to the mid-tier on a weekly cadence.

What to Watch

First, independent evaluations of 6.1 against Sonnet 5.5 at the same $2/$10 price, because that is now the default purchase decision for agent builders and the comparison nobody has published yet. Second, cached-input economics: $0.10 per million cached tokens is the quiet price war, and cache-heavy agent workloads are where it compounds hardest. Third, whether Anthropic’s counter is a Sonnet 5.5 capability bump or a price cut, and whether the weekly-cadence releases hold once the price war settles. The month’s lesson stands and accelerates: capability differences now last days, releases are now weekly, and the only durable moat on display is distribution, which is exactly what Dots is for.