GPT-6 Sol Is Half the Price of Claude Opus 5.5 on Every Line but the One Agents Read Most

Anthropic and OpenAI both launched cheaper frontier models on 22 September, and both led with a percentage. Anthropic’s Opus 5.5 page says the new model “costs 40% less to run than Opus 5.” OpenAI’s GPT-6 Sol and Luna post says it is “reducing API prices for Sol and Luna by 50%.” Both numbers are true on their own terms. Neither can be checked without a number the pages leave to the reader. Line up the two rate cards and one detail stands out. GPT-6 Sol costs exactly half what Opus 5.5 costs for input, output and cache writes. For cache reads the two charge the same: $0.20 per million tokens.

For a chat app that detail hardly matters. For a coding agent it is most of the bill.

Claude Opus 5.5 and GPT-6 Sol list prices, line by line Slopegraph on a log scale of standard-tier prices per million tokens. Output falls from $20 on Claude Opus 5.5 to $10 on GPT-6 Sol, cache writes from $5 to $2.50, and uncached input from $4 to $2, each exactly half. Cache reads are $0.20 on both, so that line is flat. Sources: Anthropic and OpenAI pricing pages, 22 September 2026. PRICE PER MILLION TOKENS · STANDARD TIER · 22 SEPT 2026 Claude Opus 5.5 GPT-6 Sol Output, Claude Opus 5.5: $20 per million tokens — Anthropic, 22 Sept 2026 Output, GPT-6 Sol: $10 per million tokens — OpenAI, 22 Sept 2026 Output $20 $10 −50% Cache write, Claude Opus 5.5: $5 per million tokens — Anthropic, 22 Sept 2026 Cache write, GPT-6 Sol: $2.50 per million tokens — OpenAI, 22 Sept 2026 Cache write $5 $2.50 −50% Input, Claude Opus 5.5: $4 per million tokens — Anthropic, 22 Sept 2026 Input, GPT-6 Sol: $2 per million tokens — OpenAI, 22 Sept 2026 Input $4 $2 −50% Cache read, Claude Opus 5.5: $0.20 per million tokens — Anthropic, 22 Sept 2026 Cache read, GPT-6 Sol: $0.20 per million tokens — OpenAI, 22 Sept 2026 Cache read $0.20 $0.20 0% Log scale, $0.10 to $40. Short-context, non-batch prices.
Every line on GPT-6 Sol's card is half of Opus 5.5's, except the one that matters most to an agent re-reading its own context. Cache reads cost twenty cents per million tokens on both.

What Anthropic’s 40% is made of

Anthropic’s own page breaks the headline apart. Input and output tokens are $4 and $20 per million, “20% less than Opus 5.” Cache reads are $0.20, “60% less than Opus 5,” and the page puts a parenthetical in the middle of that sentence: cache reads “make up the majority of agentic and coding work costs.” Cache writes drop from $6.25 to $5, which is also 20%. So the rate card on its own cuts a bill by somewhere between 20% (no caching at all) and 60% (nothing but cache reads). Where in that range a particular bill lands depends on one variable, the share of the old bill that went to cache reads.

The arithmetic is short. If cache reads were a fraction s of an Opus 5 bill, the same tokens cost 20% + 40% × s less on Opus 5.5. The headline 40% falls out at s = one half, exactly. Anthropic’s parenthetical says the share is more than a half for agentic work, which on the rate card alone would put the saving above 40%.

The headline doesn’t rest on the rate card alone, though. Further down, the page says Opus 5.5 “costs less per token than Opus 5 and uses fewer tokens per task, which nets out to a 40% drop in costs.” The earlier headline sentence frames the 40% as what Anthropic’s tests show “at default settings” on “typical workloads.” Two effects are folded into that one number, and the page doesn’t say how much each contributes or what a typical workload contains. So the 40% can’t be moved onto your own bill without knowing your own s.

A measured mix

The best public answer I can find for what s looks like in real coding-agent work is TraceLab, a University of Washington trace posted in June. It covers about 4,300 sessions, roughly 350,000 model calls and 430,000 tool calls, from 43 developers’ day-to-day use of Claude Code and Codex over about eight months. Weighted by tokens, 95.7% of input was served from the prefix cache. The median model call carried about 119,000 tokens of cached history, 875 fresh tokens and 214 tokens of output.

TraceLab then priced the trace. Cached prefix tokens came to 59.5% of total cost, fresh input (including Claude’s cache-write surcharge) to 29.2%, and output to 11.2%. The authors’ summary: “for coding agents the cost is overwhelmingly in re-reading context.” The Claude rates in their pricing table, for Opus 4.6 through 4.8, are $5, $6.25, $0.50 and $25: identical line for line to Opus 5’s. So their mix drops straight into the formula. It is approximate, since the trace blends Claude and Codex sessions, each billed at its own provider’s prices. With s = 59.5%, the Opus 5.5 rate card alone cuts the bill by 43.8%.

On this evidence Anthropic’s headline is, if anything, conservative for coding agents, even before any token-efficiency gain is counted. For workloads that cache less (chat, one-shot summarisation, anything with a changing prompt prefix) the same rate card gives closer to 20%.

What OpenAI’s 50% is made of

OpenAI’s number is the simpler of the two. GPT-6 Sol costs $2 input and $10 output. OpenAI’s pricing page lists $0.20 for cached input and $2.50 for cache writes. Against GPT-5.6 Sol’s $4, $0.40, $5 and $20, every line is cut exactly in half, so the 50% holds at any mix of tokens. It is the one headline of the two that doesn’t care about s.

The denominator is the unusual part. The cut is measured “compared with their GPT-5.6 promotional pricing,” not list. OpenAI’s GPT-5.6 Sol model page describes that promotion as “a 20% reduction in input pricing and a 33% reduction in output pricing” and says it runs “at least through November 21, 2026.” Back out the list prices ($5 and $30) and Sol’s cut against list is 60% on input and 67% on output. I’d have expected a launch page to pick the comparison that flatters the cut. OpenAI picked the one that shrinks it, presumably because the promotional price is the one customers are paying today.

Head to head, the cache line decides

The comparison neither launch page makes is Sol against Opus 5.5. The two models shipped the same day, so Anthropic’s benchmark table compares against GPT-5.6 Sol and OpenAI’s compares against Opus 5. Per token, the answer follows from the rate cards. Sol is half price on everything except cache reads, where it is the same price. Its advantage runs from 50% for a workload that never caches to zero for one that does nothing else.

How much each price cut is worth, by the cache-read share of the bill Two curves against the share of an Opus 5 bill that is cache reads, from 0 to 100 percent. Opus 5.5's per-token saving over Opus 5 rises in a straight line from 20 percent to 60 percent, reaching Anthropic's headline 40 percent exactly when cache reads are half the bill. GPT-6 Sol's per-token saving over Opus 5.5 falls from 50 percent to zero. The curves cross near 42 percent, at about 37 percent. At the 59.5 percent cache-read share measured in the TraceLab coding-agent trace, Opus 5.5 saves 43.8 percent over Opus 5 and Sol saves 28.8 percent over Opus 5.5. Writer's arithmetic on the published rate cards, holding token counts fixed. PER-TOKEN SAVING · SAME TOKENS ON BOTH CARDS · WRITER'S ARITHMETIC 0% 20% 40% 60% 0% 25% 50% 75% 100% Cache reads as a share of the Opus 5 bill Opus 5.5 vs Opus 5 GPT-6 Sol vs Opus 5.5 (dashed) 40% at a 50% cache-read share — writer's arithmetic on Anthropic's rate card Anthropic's 40% 43.8% at TraceLab's 59.5% cache-read share — writer's arithmetic 43.8% 28.8% at TraceLab's 59.5% cache-read share — writer's arithmetic 28.8% TraceLab mix, 59.5% half the bill Same token counts assumed on every card; real tasks differ by model, tokenizer and effort. Standard-tier list prices.
The two headline cuts move in opposite directions as caching takes over the bill. Anthropic's 40% is exactly what its rate card gives when cache reads are half of an Opus 5 bill; at the mix TraceLab measured in real coding-agent sessions, Sol's half-price card is about 29% cheaper than Opus 5.5.

At TraceLab’s mix Sol’s card is 28.8% cheaper than Opus 5.5’s, not 50%. The two curves in the figure cross at a cache-read share near 42%. Above that point, Anthropic’s cut from its own previous model is worth more than OpenAI’s lead over Anthropic. Re-billing the same measured tokens on all three cards shows why. The cache-read block costs the same on both new models, and Sol’s lead comes entirely from the fresh input and output around it.

One coding-agent token mix billed on three rate cards Stacked bars, indexed so the Opus 5 bill is about 100. TraceLab measured that cache reads were 59.5 percent of cost, fresh input and cache writes 29.2 percent, and output 11.2 percent, at prices equal to Opus 5's. Rebilled on Opus 5.5 the same tokens cost 56.1: cache reads 23.8, fresh input and writes 23.4, output 9.0. On GPT-6 Sol they cost 40.0: cache reads 23.8, fresh input and writes 11.7, output 4.5. The cache-read segment is identical on Opus 5.5 and Sol. Writer's arithmetic. SAME TOKENS, THREE CARDS · TRACELAB COST MIX · OPUS 5 ≈ 100 Claude Opus 5 Cache reads, Claude Opus 5: 59.5 — writer's arithmetic on TraceLab's mix 59.5 Fresh input and cache writes, Claude Opus 5: 29.2 — writer's arithmetic on TraceLab's mix 29.2 Output, Claude Opus 5: 11.2 — writer's arithmetic on TraceLab's mix 99.9 Claude Opus 5.5 Cache reads, Claude Opus 5.5: 23.8 — writer's arithmetic on TraceLab's mix 23.8 Fresh input and cache writes, Claude Opus 5.5: 23.4 — writer's arithmetic on TraceLab's mix 23.4 Output, Claude Opus 5.5: 9.0 — writer's arithmetic on TraceLab's mix 56.1 GPT-6 Sol Cache reads, GPT-6 Sol: 23.8 — writer's arithmetic on TraceLab's mix 23.8 Fresh input and cache writes, GPT-6 Sol: 11.7 — writer's arithmetic on TraceLab's mix 11.7 Output, GPT-6 Sol: 4.5 — writer's arithmetic on TraceLab's mix 40.0 Cache reads Fresh input and cache writes Output Shares round to 99.9 and labels are rounded. Token counts held fixed; each card's line-by-line ratio applied.
Rebill the same measured tokens and the cache-read block is identical on the two new models. Sol halves everything around it, and that is where its lead over Opus 5.5 comes from.

All of that holds token counts fixed, which is the only thing a rate card can do. Real tasks don’t. OpenAI’s AutomationBench table puts GPT-6 Sol at xhigh effort at 33.2% and $0.27 per task, and Claude Opus 5 at max effort at 26.9% and “11.1x” Sol’s cost per task, about $3.00. Opus 5’s card is exactly 2.5 times Sol’s on every line (5 against 2, 6.25 against 2.50, 0.50 against 0.20, 25 against 10). If that $3.00 is right, the remaining factor of about 4.4 is volume: Opus 5 spent roughly 4.4 times as much per task, in tokens weighted by price. That is my arithmetic on OpenAI’s figures.

Anthropic’s page complicates it. Its AutomationBench chart shows the same 26.9% max-effort Opus 5 result, taken from Zapier’s public leaderboard, at $1.27 per task. On that figure the volume factor is under 2. Two launch pages published the same day give the same score two costs, 2.4 times apart, and from the pages alone I can’t tell which is right. OpenAI’s table has a second caveat in its own footnote: its Claude Fable 5.1 cost “omits the cost of the Opus 5 fallbacks, which occurred on ~40% of tasks.” Anthropic’s chart puts Opus 5.5 at max effort at 40.0% for $1.37 per task, from Zapier’s own early-access run, but it compares against GPT-5.6 Sol, not GPT-6 Sol. The two companies’ tables can’t be joined. Neither company publishes the number that would settle the head-to-head: tokens per task for Opus 5.5 and Sol on identical work.

(I made a similar point about the Andon Labs stores, whose model bills outran their tills. Per-token price is the part of an agent’s cost that’s easy to read, and it’s rarely the part that decides.)

Twenty cents, and how we got here

The cache line was not always the interesting one. When OpenAI introduced prompt caching on 1 October 2024, cached input was billed at half price. GPT-4o went from $2.50 to $1.25. The discount was applied automatically and the post listed no charge for writing to the cache. Anthropic’s prompt-caching announcement, live by December 2024 according to its own update note, set a different template: writing to the cache “costs 25% more than our base input token price,” and reading it costs “only 10%.” Today OpenAI’s GPT-6 table carries a cache-write column at exactly 1.25 times input and a cached-input column at exactly one tenth of it. Anthropic has since gone further, and Opus 5.5 reads cost one twentieth of input. In two years the cheapest line on these rate cards has become the one that decides most agent bills, and on the two models launched on 22 September it costs exactly the same.

The same afternoon OpenAI published a separate post on caching for GPT-6: discounts for shared prefixes “reused within a 30-minute window,” a dashboard of hit rates, and a diagnostics tool that explains misses. The main Sol and Luna launch post adds a GitHub figure: over several months, OpenAI’s improvements “reduced the share of prompt tokens requiring fresh processing by more than 50%” for Copilot. That is a vendor-reported figure about its own customer, and it suggests which variable OpenAI thinks customers should be tuning.

TraceLab explains why the window matters. Its hit rate is 97.5% on steps triggered by a tool result and 84.4% on steps triggered by a person typing, because “the delay due to human reading, thinking, and typing can frequently exceed the prefix cache eviction time.” Across the trace, cache misses caused 3.8 times as much prefill as genuinely new input did. The cache-read share of a bill, the number both headlines depend on, turns out to depend partly on how long the developer stops to think.

An independent evaluation from January, Lumer et al., found that prompt caching cut API costs by 41% to 80% across OpenAI, Anthropic and Google on a multi-turn agentic benchmark. What went into the cache mattered: keeping dynamic tool results out of it gave steadier gains than caching everything, which could paradoxically increase latency. That fits my earlier piece on Jev’s speed multiples, where the headline likewise changed with the comparison.

The number to bring to either page

Both headlines are honest arithmetic on bases the pages mostly state. Anthropic’s 40% combines a rate cut that depends on your caching with a token saving that depends on your tasks. OpenAI’s 50% is a flat halving against a promotional price, and it says nothing about Sol against the model Anthropic shipped the same day. Against that model the cut depends almost entirely on how often Sol would be reading its own context back.

For anyone running agents, the useful number is the share of last month’s bill that went to cache reads, and the usage data will show it. Above about 42%, switching from Opus 5 to Opus 5.5 saves more per token than switching from Opus 5.5 to Sol. Below it, the reverse. Neither figure says anything about whether the work comes back right, which is a separate calculation and the more expensive one to get wrong.

References

  1. Anthropic (2026). Introducing Claude Opus 5.5. 22 September 2026. Source of the 40%, 20% and 60% claims, the “majority of agentic and coding work costs” parenthetical, the “nets out to a 40% drop” sentence, the Opus 5.5 and Opus 5 rate cards, and the AutomationBench chart (Opus 5.5 at 40.0% and $1.37 per task; Opus 5 at 26.9% and $1.27 per task) and its footnote.
  2. OpenAI (2026). Introducing GPT-6 Sol and Luna. 22 September 2026. Source of the 50% claim and its promotional-pricing basis, the AutomationBench table and Fable 5.1 fallback footnote, and the GitHub Copilot caching figure.
  3. OpenAI (2026). Pricing. OpenAI API documentation, read 22 September 2026. Source of GPT-6 Sol cached-input and cache-write prices, GPT-5.6 Sol promotional prices, and the promotion’s end date.
  4. OpenAI (2026). GPT-5.6 Sol. OpenAI API model page, read 22 September 2026. Source of the 20% / 33% reduction wording.
  5. OpenAI (2026). Better prompt caching for GPT-6. 22 September 2026.
  6. Zhu, K., Jacob, M., Ma, C., Pan, Y., Wang, S., Krishnamurthy, A. & Kasikci, B. (2026). TraceLab: Characterizing Coding Agent Workloads for LLM Serving. arXiv:2606.30560, 29 June 2026. Source of the 95.7%, 97.5% and 84.4% hit rates, the 59.5% / 29.2% / 11.2% cost shares, the median call composition, the 3.8× miss figure and the pricing snapshot.
  7. Lumer, E., Nizar, F., Jangiti, A., Frank, K., Gulati, A., Phadate, M. & Subbiah, V. K. (2026). Don’t Break the Cache: An Evaluation of Prompt Caching for Long-Horizon Agentic Tasks. arXiv:2601.06007, 9 January 2026, revised 31 January 2026.
  8. OpenAI (2024). Prompt Caching in the API. 1 October 2024. Historical background.
  9. Anthropic (2024). Prompt caching with Claude. Update note dated 17 December 2024. Historical background.