Artificial Analysis measured Anthropic's Fable 5.1 at 20% more per task than Fable 5, then 13% cheaper a week later on a new index version. List prices never changed. Our call: what really splits Anthropic and OpenAI is who ends up with the token savings, not who cuts prices

Key takeaways

In the early hours of September 17 this publication recorded a judgment of its own that was still under verification. This piece calls it "the call". The call said that in the frontier-model cost war, the variable that decides cost had moved from the list price to how many tokens a task consumes, and that Anthropic and OpenAI went opposite ways in the first week of September. Anthropic's Fable 5.1 produces about 1.7 times the output tokens of Fable 5. That buys four more points on Artificial Analysis's composite Intelligence Index. Artificial Analysis is the evaluation firm that runs one benchmark suite across many vendors' models and publishes cost per task alongside the score. Anthropic's price cut landed on cache reads alone, and list prices stayed put. Artificial Analysis still measured each Fable 5.1 task at 20% more than a Fable 5 task. OpenAI positioned GPT-6 Astra as fewer tokens and fewer retries. The call advised buyers to compare on cost per task, and warned that one vendor's price cut could move agent workloads and ordinary API calls in opposite directions. The call also recorded why it was being held back: on the OpenAI side, only the vendor's own numbers existed.

This piece takes that call apart and re-tests it. The result has three layers.

First, "20% more per task" is a real measurement, but it is tied to a benchmark that changed three times in one week. The September 1 evaluation article gave US$3.76 per task, 20% above Fable 5. On September 7 the same firm switched to a new index version, and the same model read US$7.63 per task against US$8.75 for Fable 5, about 13% cheaper. Neither vendor's list price moved by a cent that week. The new version also published two numbers from one measurement. The total spent to run the whole index was 18% higher for Fable 5.1 than for Fable 5. Cost per task, computed with the firm's own weighting, was 13% lower. Change the denominator and the sign flips.

Second, the split is real, and the OpenAI side now has an independent reading. On September 9 Artificial Analysis measured Astra at about 27,000 output tokens per task against about 78,000 for Fable 5.1, with the two tied on the composite index. Astra's cost per task came out at about 40% of Fable 5.1's. That independent reading arrived eight days before the call was written, and it pointed the way the call expected. What the call missed is the other half of the split: OpenAI set Astra's list price at 2.5 times that of its predecessor, GPT-5.6 Sol. The same firm's readings put Astra about 60% more expensive per task than Sol on the composite index and about 15% more expensive on the Coding Agent Index. Token efficiency did not reach buyers as a lower price per task. The list price absorbed it. Anthropic ran the mirror image: list price unchanged, more tokens, and part of the increase handed back through the cache-read cut. That cut applies only to Fable 5.1 and Mythos 5.1, so Fable 5.1's cache reads now cost half what Opus 5's do. Opus 5 is Anthropic's next tier down, with a list price half of Fable 5.1's. Cognition, the company behind the coding agent Devin, was the first customer to announce a move, and it moved work from Opus 5 to Fable 5.1.

Third, the number for a buyer to compare is therefore not "cost per task" but two lines on the buyer's own bill. Cost per task is not a property of a model. Five things jointly determine it: the model, the task mix, the effort setting, the tool environment and the billing channel. Whether moving from Fable 5 to Fable 5.1 saves or costs money turns on the ratio between the cache-read line and the output line on the current bill. If cache reads come to less than roughly 20% of the output line, the upgrade almost certainly costs more. If they come to more than 1.2 times the output line, it almost certainly saves money. In between, the effort setting decides how many extra tokens come out. The call's line that "the same price cut can cut both ways" needs rewriting. A price cut by itself makes nobody's bill larger. What can flip the sign is a generation change, and the most likely dividing line runs between token-billed API usage and the usage quotas of subscription plans.

The piece runs in six steps:

1. Check the number: which benchmark version measured "20% more", and what one measurement looks like under a different denominator.

2. The split holds: the independent reading on the OpenAI side, and how many extra tokens the Anthropic side really produces.

3. The other half of the split: who each vendor handed the saved tokens to.

4. The cache-read cut: whose traffic it hit first.

5. What a buyer compares: two lines on the bill, and how much of the cross-vendor gap the cache price difference eats.

6. How this publication will read such reports from now on, and what would prove this piece wrong.

This call has not run in the daily brief, and this piece is its first appearance before readers. The piece takes apart the denominator behind "20% more per task" and checks the claim that OpenAI's fewer-tokens route saves buyers money. If, within the observation window this publication has set, OpenAI brings Astra's cost per task below its own predecessor's on the same independent yardstick, the reading "token efficiency was absorbed by the list price" is wrong.

The tree is this publication's map of one question: whether frontier-model pricing power lasts. This piece works at the top of it. In July, three rival vendors attacked Anthropic's premium with the pitch "same capability at a fraction of the price". By September, OpenAI's flagship list price had instead climbed to exactly Anthropic's number, US$10 per million input tokens and US$50 per million output tokens. The list-price front has stopped at the flagship tier. The fight has moved into how many tokens a task consumes and how cache is billed.

If this is the only part read: the ratio of the cache-read line to the output line on last month's bill says more about whether a model upgrade gets cheaper or more expensive than any benchmark's cost per task.

One: check the number first. Which benchmark version measured "20% more"

The call's load-bearing reading comes from the Fable 5.1 evaluation article Artificial Analysis published on September 1 (the article; the firm's post). The article says that at maximum effort Fable 5.1 costs US$3.76 per index task, 20% more than Fable 5, because Fable 5.1 uses about 1.7 times the output tokens. The article says the cache-read price cut saves about US$1.40 per task, concentrated in the agent-style evaluations, and that without the cut the figure would be about US$5.16. Those sentences were checked word for word against the page. The same article states that Artificial Analysis supported Anthropic with pre-release evaluation. That relationship belongs next to every Fable 5.1 reading the firm publishes.

The reading is not wrong. The problem is the yardstick, which changed three times in the following week.

The figures in the first and last rows come from Artificial Analysis's own pages (Fable 5.1 model page, Fable 5 model page, v4.3 release note). The middle two rows come from the version-by-version table at TokenCost, a site that tabulates benchmark cost figures, and from TechTimes, a tech news site, which reported the three revision dates. Which version the September 1 row belongs to is a point on which the two secondary accounts disagree, so this piece writes only "launch-day version".

Three things follow from the table.

First, cost per task doubled in five days while list prices stood still. TokenCost's words: "Claude Fable 5.1 doubled in five days. $3.69 per task on September 3, $7.63 on September 8, on an unbroken $10.00 and $50.00 card." The reason is that the new version's tasks are longer and more agent-like, so each one draws more tokens. A model's cost per task is first a property of the task set and only second a property of the model.

Second, whether the new generation is more expensive or cheaper flips with the version. Launch-day version: Fable 5.1 is 20% above Fable 5. Version 4.3: US$7.63 per task against US$8.75, about 13% below. Over the same period the composite score gap narrowed from 4 points to 3, and Astra went from 5 points behind Fable 5.1 to level.

Third, and the least intuitive: the one v4.3 measurement published two numbers, and the sign flips with the denominator. The Artificial Analysis model pages carry both. One is the total spent to run the full index: US$13,128.86 for Fable 5.1, US$11,160.86 for Fable 5, 18% more. The other is cost per task, which the page describes as the total divided by the number of tasks and then weighted by each evaluation's share of the index: US$7.63 against US$8.75, 13% less.

Why do the two numbers point opposite ways? Breaking the total down shows it. In Fable 5's bill, cache reads are 28%. In Fable 5.1's bill, with the cache-read price cut to a quarter, that line is 11%, and the output line rises from 54% to 70%. Evaluations with many short tasks read little cache, so the extra output turns straight into extra spend. Agent-style evaluations with few long tasks and a heavy weight in the index read a lot of cache, and there the cut saves more than the extra output costs. Total spend weights every dollar equally and is pulled toward "more expensive" by the first group. Weighted cost per task weights every evaluation equally and is pulled toward "cheaper" by the second group. Our reading: this mechanism is inferred from the published breakdown. Artificial Analysis has not published cost per task for each evaluation, so the mechanism can't be checked evaluation by evaluation. The page also doesn't state how many times each evaluation was run for each model; if the counts differ, the two totals aren't directly comparable. The opposite signs themselves, though, are on the page in those two numbers.

An independent third party summed this up cleanly. Firecrawl, a developer-tools company, ran 57 API-billed tests of its own on September 7 (Firecrawl's write-up). Its summary table sets Anthropic's claim, 25% cheaper for typical work and up to 45% for heavy agent work, next to Artificial Analysis's reading that each task costs more. The table marks the first "True for cache-heavy work" and the second "True for cache-light work". Its verdict line: "Both claims rest on the same rate card."

So the call's first key number stands, with a narrower meaning: it is a real measurement, but it answers "what did it cost to run the September 1 task set at maximum effort", not "how much more will the upgrade cost".

Two: the split holds. The OpenAI side now has an independent reading

The call's reason for holding back was explicit. OpenAI's "fewer tokens" was self-reported, and if Artificial Analysis measured Astra's tokens per task going the other way, the split would not stand.

That reading arrived on September 9, eight days before the call was written, and in the same direction.

Artificial Analysis's Astra evaluation article says that on composite index v4.3, Astra at maximum effort scores 53, level with Fable 5.1. "At max effort, Astra uses 27k output tokens per task, about a third of Claude Fable 5.1 (max with fallback) at 78k, for the same score." Cost per task is US$3.26 against US$7.63, about 40%. On the firm's separate Coding Agent Index the two also tie, and Astra's cost per task is about 60% of Fable 5.1's, which the article attributes to "the lowest token use of any agent in the Index".

The second independent source is the public leaderboard of Terminal-Bench 4.0, a benchmark of coding tasks carried out in a command-line terminal. Harbor, the evaluation framework that runs the benchmark, maintains the leaderboard. As relayed on September 11 by CodingFleet, a site that tabulates coding-model benchmark results: Astra in OpenAI's Codex environment solves 58.2% and Fable 5.1 solves 57.9%. In CodingFleet's words, "Astra needed only 1.5B tokens ($3.3k eval cost) compared to 2.7B tokens ($6.2k) for Fable 5.1", about 47% cheaper. The Harbor leaderboard page itself hasn't been reviewed; these figures rest on CodingFleet's relay.

Set against OpenAI's own account: the September 9 enterprise launch post says that on the same Terminal-Bench 4.0, Astra's estimated API cost per task is about 63% below Fable 5.1's, and the post does not publish how that was calculated. Harbor's reading on this benchmark is about 47% cheaper, the same direction as OpenAI's self-report, and a notch smaller than 63%. Artificial Analysis's figure of about 57% cheaper comes from a different yardstick, the composite index, so it corroborates the direction and doesn't count as a second reading of the same benchmark.

On the Anthropic side, "more tokens" also has independent readings, but the size narrows.

• Firecrawl's 57 runs: at the same effort level, Fable 5.1's output tokens are 1.37 times Fable 5's at low effort, 1.12 times at high and 1.30 times at maximum. One detail Firecrawl measured: at maximum effort the extra tokens are all reasoning tokens, and the visible text is 6% shorter. Reasoning tokens are invisible to the user but billed as output tokens.
• Anthropic's own migration guide lists two behaviour changes that cost more tokens. When editing files, the model "is more likely to rewrite the entire file than make a targeted edit", and "the rewrite costs more output tokens and time". In agent loops, Fable 5.1 "may issue one tool call per turn" where Fable 5 would send several in one turn, and "the extra turns cost tokens". The guide supplies prompt wording for each.
• Readings in the other direction exist. Cognition, the company behind the coding agent Devin, says that on its own coding benchmark "Fable 5.1 takes 33% fewer tokens to complete the same tasks relative to Opus 5". Three customers quoted in Anthropic's launch post report using fewer tokens. These are comparisons against Opus 5 or against specific workloads, not against Fable 5 at the same effort level, and all of them come from partners presented on launch day.

So "Anthropic spends more tokens for a higher score" narrows to this: against Anthropic's own predecessor at the same effort level, 10% to 40% more tokens; the 1.7 times figure is the launch-day version's reading at maximum effort.

One contrary reading remains to be dealt with. MindStudio, an AI tools company, wrote on September 6 that on the same set of coding tests Astra cost about US$198 and Fable 5.1 about US$113, so Astra was 75% more expensive. But MindStudio ran Astra in Codex's highest reasoning tier, and OpenAI's own description of that tier, added in July, is "coordinating four agents in parallel by default, trading higher token use for stronger results" (as relayed by Latent Space). So MindStudio's figure leaves the split standing. What it shows is that "fewer tokens" means something only under the same effort label, and the labels don't line up across vendors. TokenCost makes the same point: "max" and "high" mean different things at different vendors.

Three: the other half of the split. Who keeps the saved tokens

The call's picture was this: Anthropic has users pay more tokens for more capability, and OpenAI uses fewer tokens to save users money. Section two confirmed the direction of the tokens. "Save users money" needs a third number, the list price.

OpenAI's flagship list prices, in order:

• GPT-5.4: US$2.50 per million input tokens, US$15 per million output tokens.
• GPT-5.5, April: raised to US$5 and US$30, the first generation-on-generation increase in the line. OpenAI said at the time that output tokens fell by about 40%, for an effective increase of about 20%.
• GPT-5.6 Sol, July: the launch price on this publication's record is US$5 and US$30 (Latent Space). Artificial Analysis on September 9 gives Sol's "current prices" as US$4 and US$20. The price change in between is not documented in this piece's sources.
• GPT-6 Astra, September: US$10 and US$50 (OpenAI pricing page). Artificial Analysis's words: "Pricing is 2.5x GPT-5.6 Sol's current prices across the board, up from $4/$20 to $10/$50". The 90% cache-read discount stays, which puts cache reads at US$1 per million.

This is the second time in six months that OpenAI has made the same move: raise the list price and say the new model uses fewer tokens.

The same Artificial Analysis article gives the result. On the composite index, Astra at maximum effort is "~60% more expensive than GPT-5.6 Sol. The higher price per token is partially offset by lower token use". On the Coding Agent Index, Astra costs US$7.09 per task, "~15% more than GPT-5.6 Sol (max) for a 7 point higher score".

The two vendors side by side:

"About 7% fewer" comes from the output tokens per task on the Artificial Analysis model pages, 27,200 against 29,300. "About a third" is the article's own wording in its coding-agent section, "one third of the tokens of GPT-5.6 Sol"; the raw token counts behind that phrase are not in this piece's sources. The two readings describe different indices, and on the composite index Astra uses only slightly fewer tokens than its predecessor. The "about a third" in this table is also a different comparison from section two's "a third", which set Astra against Fable 5.1.

Read across the table, neither vendor gives buyers anything clearly cheaper in the row "cost per task against own predecessor". The difference is the knob each vendor turned: OpenAI used token efficiency to lift the list price; Anthropic absorbed token growth with the cache price. The saving buyers actually get is cross-vendor. At the same list price, Astra's cost per task is 40% to 60% of Fable 5.1's. That gap is between OpenAI and Anthropic. Against its own predecessor, OpenAI offers no such saving. On the OpenAI side this piece has list prices and benchmark readings; a buyer's public account of why Astra's premium is worth paying is absent from its sources, so that half of the evidence is missing.

This also answers this publication's July call. At that time three vendors attacked Anthropic's premium with the line "same capability at a fraction of the price", and OpenAI competed with cheaper product tiers. Two months later OpenAI's flagship list price has risen to match Anthropic's flagship to the cent. At the flagship tier, OpenAI no longer competes on list price. The competition has moved to tokens per task.

Four: the cache-read cut. Whose traffic it hit first

The call read Anthropic's cache cut as "a targeted subsidy for agent workloads rather than a list-price response to rivals' attacks". The first half mostly holds. The target needs a more precise name.

Start with where the cut lands. The footnote on Anthropic's official pricing page: "Cache hits and refreshes on Claude Fable 5.1 and Claude Mythos 5.1 are priced at 0.025x the base input price. All other models use the standard 0.1x multiplier." Cache-write prices did not move: US$12.50 per million tokens for a five-minute cache, US$20 for a one-hour cache.

Limiting the cut to these two models means one thing. Fable 5.1's cache-read price, US$0.25 per million tokens, is now half of the US$0.50 charged by Opus 5, Anthropic's next tier down. On list price Opus 5 is half of Fable 5.1. On cache reads the order inverts.

The first to say so publicly was Cognition. Anthropic's launch post quotes Cognition: "We're moving our Opus 5 traffic in Devin to Claude Fable 5.1 on launch day. … with the new cache read pricing a Fable-class model is finally economical for the workloads we'd kept on Opus, starting with code review." Cognition's own blog gives numbers: on its coding benchmark at medium effort, Fable 5.1 costs US$2.68 per task, Opus 5 US$3.51 and Fable 5 US$5.84. Cognition's explanation is that at the old cache-read price most of the money in a Fable 5.1 task went to re-reading content already seen, and after the cut the same task fell from about US$5 to US$2.68. And on Opus 5: "Opus 5 reads its cache at twice Fable's new rate, so despite the cheaper sticker it lands above Fable."

So the first effect this piece can see is a move inside Anthropic's own product line: cache-heavy work moving up from the Opus tier to the Fable tier. That doesn't rule out Anthropic also answering OpenAI. This piece doesn't rule on motive; it records who moved first and where from. Cognition is a launch-day partner, and its numbers are self-reported.

Two more things travel with this cut.

First, the next line of the cache bill didn't move. After the cut, in Artificial Analysis's Fable 5.1 bill, cache writes are 16%, already larger than cache reads at 11%.

Second, the cut lands only where usage is billed by token. Anthropic's launch post says the cut applies "wherever usage is billed by token, such as on our API". How subscription plans count cache against their usage quotas, the post doesn't say. A September 3 post on paddo.dev, a developer's blog, puts the point in its title: "Two Meters: Claude Fable 5.1 Is Cheaper on the API and Hungrier on Max" (the post). The author's own Claude Code logs show 1,207 output tokens per turn against 990 for Fable 5, 22% more, with cache reads per turn about the same. On the API the cut absorbs the extra output. On a subscription quota nothing absorbs it. The evidence here is user self-reports and one author's personal logs, and whether quota metering was adjusted is, in the author's own words, "the open question". But if it holds, the sign flip the call predicted most likely falls not between agent workloads and ordinary calls but between the API and subscription plans.

Five: what a buyer compares. Two lines on the bill

Section one said cost per task is a property of the task set. Section four said it is also a property of the billing channel. So what does a buyer actually compare?

Fable 5 to Fable 5.1: the ratio of cache reads to output

Take a Fable 5 bill where the output line is O and the cache-read line is C. After the move to 5.1, output becomes O times r, where r is the multiple of extra tokens. The cache-read price drops to a quarter, but the number of reads may rise by a factor c as turns multiply, so that line becomes C × c × ¼. The new bill is cheaper than the old one when C exceeds O × (r − 1) ÷ (1 − c/4) (our calculation).

With numbers:

• 12% more output and 12% more reads: the cache-read line only has to exceed 17% of the output line for the upgrade to save money.
• 30% more of each: the threshold is 44%.
• 70% more of each, which is the launch-day version at maximum effort: the threshold is 1.22 times.

So the rough rule is this. Cache reads below 20% of output, and the upgrade is almost certainly more expensive. Cache reads above 1.2 times output, and it is almost certainly cheaper. In between, the effort setting decides how many extra tokens come out. The formula ignores cache writes and uncached input, and the former grows with the number of turns, so the real threshold sits a little higher. Checked against the Fable 5 bill in Artificial Analysis v4.3: cache reads are 52% of output. On the weighted per-task token multiple of about 1.17, that sits above the threshold, a saving. On the total-spend output multiple of about 1.5, it sits below, a cost. Both match the two opposite-signed numbers from section one.

The formula is this publication's own arithmetic, not any party's claim. Its use is this: no benchmark needs waiting for. Last month's bill already holds both numbers.

Fable 5.1 against Astra: how much the fourfold cache gap buys

At the same list price, Anthropic's cache-read price is a quarter of OpenAI's. Can that gap offset Astra's fewer tokens?

Artificial Analysis v4.3 publishes both vendors' cache-read spend for the full composite index. Dividing that spend by each vendor's official cache-read rate, US$0.25 and US$1 per million tokens, gives the read volume. Fable 5.1 read about 5.5 billion cached tokens and Astra about 1.1 billion, 5.2 times as many, for only 1.3 times the money (our calculation). The fourfold price gap is almost entirely eaten by Fable 5.1's extra reads, because every extra turn the model runs re-reads the context. The total bill for Fable 5.1 is 2.47 times Astra's; at Astra's cache price it would be 3.25 times. The cache cut narrows the gap, but the bulk of the gap is output. These read volumes are this publication's inference, not figures Artificial Analysis publishes. The inference leaves out the long-context surcharge, and if the benchmark triggered it, Astra's read volume is overstated.

Three cross-vendor readings in a row show the gap shrinking as tasks get more agent-like: about 2.3 times per task on the composite index, about 1.9 times in Terminal-Bench 4.0 run cost, about 1.7 times on the Coding Agent Index. The tool environments differ, and the latter two are each vendor's own agent environment, so this piece doesn't attribute the narrowing to the cache price gap alone. It records the direction, not a rule.

Three numbers a vendor would have to supply

1. The effort level and tool environment behind any quoted tokens-per-task figure.

2. The share of a given workload's bill that goes to cache reads and to cache writes.

3. How cache reads count against a subscription plan's quota.

No vendor has publicly answered the third.

"Only one third party measures this" has to be withdrawn

The call wrote that only Artificial Analysis, among third parties, publicly measures tokens per task. That doesn't hold. This piece found at least five third parties or users who publish cost per task or token use. Vals AI, an evaluation platform that runs its own test sets, lists Fable 5.1 at US$28.92 per test on its model page. The Harbor-maintained Terminal-Bench 4.0 leaderboard has cost and token columns. Cognition publishes cost per task on its own coding benchmark. Firecrawl published its 57 runs. Databricks published in July a cost-per-completed-task benchmark on its own codebases. The Vals AI reading the call itself cited, on a task set supplied by Anthropic, is itself a cost per task.

The defensible narrowed version: the most-cited yardstick is Artificial Analysis, and that yardstick changed version three times in a week, each time changing the cost-per-task number, and once changing its sign.

Six: how to read "cost per task" from here

Our call

Frontier flagship list prices converged in September on one number, US$10 per million input tokens and US$50 per million output tokens; the list-price front has stopped at the flagship tier, and the cost war has moved into how many tokens a task consumes and how cache is billed. Anthropic's Fable 5.1 and OpenAI's GPT-6 Astra did go opposite ways on tokens per task. Each direction is backed by two independent sources. But neither vendor turned the change into a cheaper per-task price against its own predecessor. OpenAI turned its token efficiency into a higher list price. On Artificial Analysis's two indices, Astra costs 15% to 60% more per task than GPT-5.6 Sol. Anthropic absorbed token growth with the cache-read cut. That cut applies only to Fable 5.1 and Mythos 5.1. Its first visible effect is inside Anthropic's own product line: cache-heavy work moving from Opus 5 up to Fable 5.1, with Cognition the first customer to announce such a move. "Cost per task" is not a property of a model; it is a property of task mix, effort setting, tool environment and billing channel, and the most-cited yardstick changed version three times in a week and flipped sign once. From here, this publication puts four questions to any report that "model X costs N% more or less per task": which benchmark version, which effort level, what share of the bill is cache reads, and whether the usage is billed through the API or a subscription.

What would prove this wrong

1. OpenAI brings Astra's cost per task below its own predecessor's. Specifically: on any version of the Artificial Analysis or Harbor benchmarks, at the same effort level, Astra's cost per task falls below GPT-5.6 Sol's. A lower list price, a lower cache price or a further drop in tokens would each count. The reading "token efficiency was absorbed by the list price" is void that day. The observation window, set by this publication, runs to December 31, 2026.

2. Anthropic extends the 0.025x cache-read rate to Opus 5. The reading "the cut hit Anthropic's own Opus traffic first" weakens, because work left on Opus no longer needs to move.

3. A third-party comparison appears with the same tool environment, the same effort level and the same task set, publishing the cache-read share, and shows Fable 5.1 cheaper than Fable 5 even on work that reads little cache. Section five's two-line rule is wrong.

4. Anthropic announces that subscription quota metering now uses the new cache ratio, or a reproducible measurement shows Max-plan quota draining more slowly on 5.1 than on 5. Section four's paragraph on the API-versus-subscription dividing line is withdrawn.

5. Artificial Analysis publishes cost per task for each evaluation, and on v4.3 the agent-style evaluations turn out no cheaper than the rest. Section one's explanation of the sign flip is wrong, and only the fact that the two numbers point opposite ways remains.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.

Sources cited in other language editions

These sources are linked in the body of this piece's other language editions but not in this one.

• magazine.sebastianraschka.com/p/gpt-6-astra-looped-transformers-and
• datacamp.com/blog/gpt-5-6-sol-luna-terra