OpenAI and Anthropic each shipped new models within about an hour of one another on 22 September 2026, and both came with lower per-token prices, according to a community write-up on dev.to that draws on reporting by Simon Willison. The post describes the result as a price war among frontier AI vendors. The account is a single community source, so the figures below are the author's and Willison's claims rather than independently verified list prices.
GPT-6 Sol and GPT-6 Luna, according to the post on OpenAI's side, carried prices about 50% below what their GPT-5.6 counterparts cost. For GPT-6 Luna, the quoted rates are $0.10 per million input tokens alongside $0.50 per million output tokens; the author calls it one of OpenAI's cheapest shipped models, with only smaller Nano-tier offerings below it. GPT-6 Sol carries quotes of $2 per million input and $10 per million output, occupying the same tier that GPT-5.6 Terra held before.
That overlap matters for anyone still paying for Terra. If Sol now sits at Terra's old price point, the author's framing is that there is effectively no reason to keep using Terra. This is an editorial judgement in the source, not a benchmark result, and it assumes Sol performs at least as well for a given workload.
The comparison looks larger still against OpenAI's forward schedule. The post states GPT-5.6 pricing was set to rise 25% in November, so the gap between the old and new tiers widens against that baseline rather than the current one. The source does not say whether that increase was cancelled, deferred or left in place for the older models.
Anthropic's move is smaller in headline terms but aimed at a different cost centre. Claude Opus 5.5 received a 20% cut versus Opus 5.0 through 4.8, moving from $5 per million input and $25 per million output to $4 and $20 respectively. That is a straightforward reduction on the same tier rather than a new cheaper tier.
The more consequential change for agent workloads is cache-read pricing, which the post says fell 60%. Willison's stated rationale is that in longer conversations more than 90% of input tokens are served from cache. If that proportion holds for a given agent, a 60% cut to cache reads hits the dominant line item, while the 20% headline cut applies to the smaller uncached remainder.
Cheaper Sonnet 5.5 and Haiku 5.5 models are coming, Anthropic also signalled, and Haiku's need to compete with GPT-6 Luna's pricing was specifically flagged by the company, per the post. Because those statements concern models not yet released, the evidence here establishes no pricing, no capability and no date.
The range is framed by a third vendor. Released the day before, Grok 4.7 now sits between the two camps: at launch it substantially undercut GPT-5.6 Sol, yet on input pricing GPT-6 Sol now roughly matches it. Grok's output pricing is absent from the post, leaving the comparison partial.
The source also reports mixed outcomes on Willison's standard benchmark for drawing a pelican. On two occasions Claude Opus 5.5, at its maximum reasoning setting, did not succeed: it hit the output ceiling of 128,000 tokens while it was still working through what was a simple SVG request. Reportedly, roughly 20 minutes were needed for each failed attempt, at a cost of $2.56 apiece. This is one person's test on a single task, not a finding about general capability.
A practical data point is Willison's own response. His default coding tools were switched to GPT-6 Sol and Claude Opus 5.5, and a public Datasette demo was upgraded to GPT-6 Luna, according to the post. What one prominent developer's tooling switch signals is perceived value, not a measurement of quality.
For freelancers and small studios, the arithmetic is the interesting part. A cheaper output token reduces the cost of long generations, while a cheaper cache read reduces the cost of repeated context in agent loops. The two cuts therefore favour different workflows, and a team running short one-shot prompts will see less benefit from Anthropic's cache change than a team running long agent sessions.
The same logic applies to model choice. If Sol now occupies Terra's old price band, the decision shifts from cost to output quality and latency, which the supplied evidence does not compare. The post offers no benchmark table, no latency figures and no error rates across the models, so a migration decision cannot be justified from this source alone.
There is also a timing caveat. The post describes same-day releases and a price war, but coincidence of dates is not evidence of coordination, and the source does not claim any. Two vendors cutting prices in the same week is consistent with competition, with independent roadmap timing, or with both.
What remains unknown is substantial. The evidence gives no confirmation from OpenAI or Anthropic, no effective dates for the new rates, no regional or tiered pricing, no rate limits, no context-window details and no statement on whether older models keep their old prices. The November increase for GPT-5.6 is described as scheduled, with no outcome reported.
The practical reading for this audience is to treat the quoted rates as a prompt to check current vendor pricing pages before renegotiating a fixed-price contract or re-estimating a token-heavy build. If the cuts hold, agentic and long-context work benefits most, and the cache-read reduction is the line worth modelling first. If they do not, the older tiers remain the safe assumption.
The clearest supported conclusion is narrow: on 22 September 2026, two frontier vendors released models within an hour of each other with lower quoted token prices, and one of them cut cache-read pricing sharply. Everything beyond that, including which model is better and whether the cuts persist, is not established by the evidence supplied.