Claude Opus 5.5 and GPT-6 Sol: flagship capability at mid-tier prices, three weeks later

On 22 September 2026, Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and GPT-6 Luna, on the same day. The two launches say the same thing in different words. The capability that was the flagship at the start of September now sells at a mid-tier price, and it took about three weeks to get there.

A note on sourcing before the numbers. Anthropic's announcement and model documentation are quoted directly below. OpenAI's launch page blocks automated readers, so the OpenAI figures here come from VentureBeat's report and from Digital Applied's line-by-line reading of OpenAI's page and pricing tables.

Three weeks from flagship to mid-tier#

Anthropic's sequence is the cleaner of the two. Claude Fable 5.1 was released on 1 September at $10 per million input tokens and $50 per million output tokens, according to Anthropic's own model documentation. Three weeks later, the Opus 5.5 announcement opens with this: "It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5."

The Opus 5.5 price is $4 input and $20 output per million tokens, which Anthropic says is 20% below Opus 5. Cache reads drop to $0.20 per million, 60% below Opus 5, and Anthropic adds that it "generates output more than 30% faster than Opus 5." The model documentation lists both Fable 5.1 and Opus 5.5 with a 1M token context window and 128K maximum output. One detail in that table is easy to miss: Fable 5.1 defaults to high effort and Opus 5.5 to medium.

OpenAI's sequence is similar in timing and different in claim. GPT-6 Astra launched on 3 September, and Digital Applied's pricing table puts it at $10 / $50. GPT-6 Sol and Luna followed 19 days later. Sol is $2 input and $10 output per million tokens. Luna is $0.10 and $0.50. VentureBeat reports OpenAI's positioning: Sol for "complex work that developers and knowledge workers perform repeatedly", Luna for "more tightly defined jobs such as summarization, extraction and answering straightforward questions."

The asymmetry matters. Anthropic claims its new mid-tier model reaches its flagship on most work. OpenAI does not claim that Sol reaches Astra, and its launch page, as quoted by Digital Applied, says "GPT-6 Astra continues to be our best model across the board." Digital Applied's reading of OpenAI's chart data is blunt: the gains are "mostly in cost, not in top scores." On two charts, the older GPT-5.6 Sol at max effort still scores higher than GPT-6 Sol: 72.7% against 68.8% on DeepSWE and 66.2% against 64.4% on OSWorld, though at more than twice the cost per task.

Read the price twice#

Headline token prices are the least useful number in either launch. Three details change the bill more.

Per-task, not per-token. Anthropic's "40% less" is a claim about typical workloads at default settings, combining the lower price with fewer tokens per task. The token price alone fell 20%. If your workload does not shrink its token count the way Anthropic's did, your saving is closer to the smaller number.

The cache line. Anthropic's own sentence on this is worth taking seriously: cache reads "make up the majority of agentic and coding work costs." For an agent that reuses a long system prompt and tool definitions on every turn, the cache-read price is the real price. Model your costs on it.

The baseline and the fine print. OpenAI's 50% cut is measured against GPT-5.6 Sol's promotional rate of $4 / $20, which Digital Applied notes OpenAI guarantees only "at least through November 21, 2026." The new GPT-6 rates are not promotional; an OpenAI spokesperson told VentureBeat they are permanent. And above 272K input tokens, Digital Applied reports that a whole Sol or Luna request is billed at twice the input rate and one and a half times the output rate. The 1.05M token window is real. Filling it costs more per token than the headline suggests.

Effort has become a cost dial too. Both vendors now ship several effort levels, and both new models default to medium. The default is a choice somebody made for you. It is not necessarily the right one for your task.

The launch tables contain answers from other models#

This is the part I would want every buyer to read slowly.

Opus 5.5 ships with safeguards on cybersecurity, biology and distillation, "all of which fall back to another model transparently." Anthropic is specific that "most cybersecurity tasks will be re-routed to Opus 4.8." Its benchmark footnote is equally specific: "When they intervened, cybersecurity tasks were completed by Claude Opus 4.8, and biology and frontier LLM development tasks were completed by Claude Opus 5."

OpenAI's footnotes, as Digital Applied reports them, show the same thing from the other side. The Fable 5.1 cost on OpenAI's AutomationBench chart leaves out the Opus 5 fallback runs, which OpenAI says happened on about 40% of tasks. Competitor results were "taken from publicly available reports," not rerun.

None of this is hidden, and both companies deserve credit for footnoting it. But it means the model name on a benchmark row, and potentially on your invoice, is not always the model that produced the answer. Anthropic says as much in its own words: "at these levels of capability we've found that benchmark margins have become a less reliable guide to real-world differences."

What this does to model routing#

Most AI products I would build today route work between models: a strong model for the hard cases and a cheap one for the volume. That routing table encodes a belief about where the capability gap sits. In September the gap moved.

If a mid-tier model now matches the flagship on most of your tasks, the flagship route should shrink to the specific task types where your own evaluations still show a difference. If Luna handles extraction and summarisation at $0.10 / $0.50, the definition of "cheap tier" has moved as well.

Three practical consequences follow.

Routing is configuration, not code. If changing which model handles which task needs a deploy, you will fall behind a release cycle that now runs in weeks.

Log which model answered. Given the fallback behaviour above, find out from your vendor how a fallback is reported in the response and usage data, and record it per request. A quality regression you cannot attribute to a model is one you cannot fix.

Switching vendors is not a price decision. Tool calling, caching and API surfaces differ. Digital Applied points out one example from OpenAI's Sol model page: Chat Completions supports function calling only when reasoning effort is set to none, so agents that call tools with reasoning switched on need the Responses API. That is a migration, not a config change.

How I would rerun the evals#

The launch tables are a reason to test, not a result. The version of the test I would run:

1. Use your own tasks. Pull a representative, anonymised set from real traffic, with a written pass criterion for each. A task set you did not choose measures somebody else's product.
2. Measure cost per completed task. Include retries, cache hit rate and the effort level. A cheaper token that needs two attempts is not cheaper.
3. Test more than one effort level. Digital Applied's reading of OpenAI's data shows the highest-scoring effort level differs by benchmark, and sometimes a higher setting scores lower.
4. Count refusals and fallbacks as outcomes. Anthropic notes that Zapier ran AutomationBench without fallback models, so safeguard interventions counted as failures. That harness choice changed the score. Your harness choices will change yours.
5. Pin versions and rerun on every release. A result from August is now a historical document.

And then decide. A comparison that ends in "both are good" has the same problem as a successful pilot that is not a decision. Write down which task types move to which model, and why.

What to watch next#

Anthropic said at launch that "Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks". Sonnet 5.5 arrived six days later, on 28 September, at $2 / $10, the same headline rate as GPT-6 Sol, and Anthropic says that in its testing "it costs up to 30% less per task than its predecessor." That is the next head-to-head worth running, and Haiku 5.5 is still to come. On OpenAI's side, the number to watch is what GPT-5.6 Sol costs after 21 November, because that decides whether the "50% cheaper" comparison still holds. And across both, I would watch for fallback behaviour becoming a first-class field in API responses and usage reports. When the model you pay for can hand your request to another model, that fact belongs in your logs.