Obsidian 3: A Hidden Benchmark Row, a 3.6× Bill, and the Fable 5.5 Question

The most interesting number on Cognition's new coding leaderboard this week belongs to a model nobody can call. A hidden entry labeled Obsidian 3 appeared in the data file behind FrontierCode 2.0 with a high-effort score of 46.77%, above Claude Opus 5.5 at 45.5% — the model Cognition's own table lists first. The row ran in the claude-code harness, which Anthropic's models alone use on that board, and the speculation split immediately between Claude Fable 5.5, the flagship Anthropic has not shipped, and a mid-cycle refresh one poster called "Opus 5.6." None of that is confirmed. Anthropic has said nothing, no model card or API ID exists, and every Obsidian 3 figure in this piece is an unverified claim from X posts rather than a result we could reproduce.

What can be checked is narrower and more useful: the public file the leak pointed to, the effort curve attributed to the hidden row, and the arithmetic behind the "3.6× cost" claim. Put those together and the story stops being "Anthropic's next model beats its current one." It becomes something more specific — a row that peaks at high effort and then goes backwards, a comparison that only works if you hold different effort levels on each side, and a data file that does not contain the row at all.

What leaked, and when

Cognition published FrontierCode 2.0 on 27 October. The benchmark is deliberately hard to game: 132 tasks built from real open-source repositories by more than 20 maintainers, each taking roughly 40 hours, graded on whether a patch would actually be merged rather than whether it passes a test. Every run is repeated three times, the tasks are private, and a single failed blocking criterion scores zero. Cognition's own table put Claude Opus 5.5 first at 45.5% in max effort, GPT-6 Astra second at 42.8%, Claude Sonnet 5.5 third at 41.0%, and no model above 50%. Artificial Analysis adopted the benchmark at 15% of its new Intelligence Index v5 and independently reported the same leader, Claude Opus 5.5 at 46% when rounded.

That evening, at 21:10 UTC on 27 October, @tensorson posted that Cognition's data file contained a model the table did not show: "Obsidian 3," at 46.8%, "above #1 Opus 5.5 (45.5%)." The post noted the claude-code harness and guessed Fable 5.5. Twelve hours later, a linux.do thread listed five separate effort scores for the row — low 39.19%, medium 39.84%, high 46.77%, xHigh 45.90%, max 45.53% — and attached a copy of the data. At 09:12 UTC on 28 October, @teortaxesTex quote-posted with the cost figures: Opus 5.5 at high scores 34.5% for $2.42 a rollout, Opus 5.5 at max scores 45.5% for $12.82, and Obsidian 3 at high scores 46.7% "apparently for $8.62," reaching no further even at $47 for max. "Is FrontierCode already broken?" the post asked. "Or is Fable 5.5 a flop?" At 16:18 UTC, @CodexResets1 amplified the same numbers and asked whether the model was Fable 5.5 or Opus 5.6.

The file's Last-Modified header reads 17:44 UTC on 28 October. The version we downloaded a day later is 32,055 bytes, lists 24 models, and contains no row called Obsidian 3 anywhere in its structure. A cache-busted request returned the same file hash. The official leaderboard table never showed the row, BenchLM's 20-row mirror did not show it either, and Anthropic has published nothing.

One row, five effort levels, three different stories

The published half of this comparison is solid, because it comes straight from Cognition's downloadable data. Claude Opus 5.5's mergeability-adjusted score rises at every step: 14.7% at low, 30.1% at medium, 34.5% at high, 43.4% at xHigh, and 45.5% at max. Its average task cost rises from $2.42 at high to $12.82 at max. That is the curve of a model that gets better when you pay it to think longer.

The claimed Obsidian 3 curve does not behave that way:

• Low effort — 39.19% claimed, already above Claude Opus 5.5's high-effort 34.5%.

• Medium effort — 39.84% claimed, a gain of 0.65 points.

• High effort — 46.77% claimed, the peak, at a claimed $8.62 per task.

• xHigh effort — 45.90% claimed, down 0.87 points.

• Max effort — 45.53% claimed, down another 0.37, at a claimed cost near $47.

Five effort settings, and the best score comes third. Spending roughly 5.5× the high-effort budget to reach max buys a score 1.24 points lower. Cognition itself warns that this can happen: higher reasoning settings sometimes reduce scores because models write more code and touch more files, and the benchmark's own examples include a max-effort run that deleted a public API class while fixing a race condition. But the size of the drop matters. A model that stalls at the top is an early checkpoint or a distillation artifact far more often than it is a finished flagship.

Now the comparisons, because the headline depends entirely on which one you pick:

• Best versus best — Obsidian 3's 46.77% at high against Claude Opus 5.5's 45.5% at max is +1.32 points, at $8.62 against $12.82, or about 33% cheaper. On this reading the hidden row wins on both axes.

• High versus high — 46.77% against 34.5% is +12.24 points, at $8.62 against $2.42, or 3.56× the cost. On this reading it is a large quality gain at a large price.

• Max versus max — 45.53% against 45.45% is +0.08 points, at roughly $47 against $12.82, or 3.67× the cost. On this reading it is a tie at a ruinous price.

All three are arithmetically true. That is why the X thread argued past itself: @tensorson compared high to max and called it a 1.3-point win; @teortaxesTex compared high to high and asked whether the benchmark was broken. The "3.6× cost at the same effort" figure is accurate, and so is "cheaper than Opus 5.5's best run" — they describe different pairs of runs.

There is a bigger caveat underneath all of it. Cognition's data stores two numbers per run: a raw correct rate and a mergeability-adjusted score. The published 45.5% for Claude Opus 5.5 is the adjusted figure; its raw correct rate at max is 49.7%, and at high it is 37.9%. The adjustment costs the public Claude rows three to four points. We do not know which column the leaked Obsidian 3 numbers came from. If 46.77% is a raw correct rate, its adjusted score would land near 43% — below Claude Opus 5.5 — and the entire "above #1" claim disappears. That single unknown is larger than the 1.3-point margin everyone argued about.

Why the claude-code label is the entire case for Anthropic

A harness is scaffolding, not a fingerprint. FrontierCode runs each model inside an agent loop, and the data file maps models to the loop they used: OpenAI's rows run in codex, Google's in devin-cli, xAI's in grok-build, Meta's Muse Wonder in muse, Cognition's own SWE models in devin-cli-swe1, and all six Claude rows in claude-code. An anonymous row sitting in the Claude harness is a strong hint that the model behind it was built to work with Anthropic's tooling. It is not proof that Anthropic built it. A benchmark operator can test a partner's pre-release model in whatever harness fits, and a lab can test its own model under an embargoed name.

The name itself tells you almost nothing. Anthropic's public lineup — checked against its own models overview on 29 October — is Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5. There is no Fable 5.5, no Opus 5.6, and no product called Obsidian. "Obsidian 3" is a gemstone-style codename of the kind labs use internally, but there is no public registry to check it against, and the trailing "3" does not map to any sequence Anthropic has shipped.

That leaves three readings, and the evidence does not choose between them:

• A Fable-tier build. Claude Fable 5.1 is the $10 / $50 flagship for demanding reasoning, launched 1 September, and the Fable line is the obvious gap in the current lineup. @tensorson's guess, and several replies to the post, landed here because the cost profile looks like a premium tier rather than a value tier.

• A mid-cycle Opus refresh. @teortaxesTex's prior is that "Obsidian 3 is Opus 5.6," on the argument that Anthropic has shipped mid-cycle models with little reasoning gain before. The flat top of the claimed curve fits that story, but the low-effort scores — 39.19% against Claude Opus 5.5's 14.7% — look like a different training run, not a small refresh.

• A distillation artifact or early checkpoint. @teortaxesTex's other hypothesis, and the one the curve shape supports best: strong at low and medium effort, peaking at high, then regressing when given more room. A finished frontier model should not get worse every time you increase its budget.

What the IPO timing adds — and what it doesn't

The signal that started this thread was a question about timing: why hasn't Fable shipped before the IPO? Anthropic publicly filed its S-1 on 26 October, the day before FrontierCode 2.0 launched, for an estimated $75 billion raise on Nasdaq under the ticker ANT. The roadshow is expected the week of 9 November and pricing the week of 16 November, according to Renaissance Capital. A next flagship that lands before pricing would give the listing a product story; one that lands after gives it a roadmap story.

But a hidden benchmark row is not a launch signal. Labs test pre-release models with benchmark operators routinely, usually under embargo, and the operator's job is to keep the model anonymous until the vendor says otherwise. If Cognition tested an Anthropic build, the correct reading is that Anthropic is still iterating — not that a release is days away. The only official data point remains Anthropic's documentation, which lists four current models and none of them Obsidian. The IPO explains why people are watching the Fable tier so closely. It does not make the row real.

What a reader can actually do today

Nothing about Obsidian 3 is callable, priced or documented, so it changes no decision this week. If you need the best mergeability score available now, Claude Opus 5.5 is public: $4 input and $20 output per million tokens, a 1M-token context window, 128K max output, API ID claude-opus-5-5, released 22 September. If cost is the constraint, GPT-6.1 Sol scored 39.0% at roughly $1.03 per task on the same board — a far better score-per-dollar trade than either Claude run, and the number worth checking before you assume the top of the table is the only useful row.

Claude Opus 5.5 is routable through OrcaRouter at Anthropic's list price with no markup, alongside more than 200 other models on one API with automatic failover. That matters for the one experiment this leak actually invites: run Claude Opus 5.5 at high and at max on your own workload and measure the cost per accepted patch. FrontierCode's own curve says the last two effort steps cost $10.40 more per task for 11 points; your repository may not price that the same way. The effort setting and the per-token price are Anthropic's own, passed through, so the comparison is a parameter change rather than a second contract — and if Anthropic cuts the list price, the cut is live on the route the same day.

What would make Obsidian 3 a fact

• The row comes back in a downloadable data file, with a model key and all five effort entries, so the numbers can be read without a screenshot.

• Cognition says whether it tested a pre-release model, and under what arrangement. A benchmark operator confirming an embargoed run would settle the provenance without naming the vendor.

• Anthropic publishes a model card, an API ID or a docs entry. That is the only event that turns a benchmark row into a product.

• A price appears. A model with no price is a benchmark entry, not an option a reader can choose.

• A second operator measures it. Artificial Analysis runs its own FrontierCode 2.0 harness with different outer settings; an Obsidian row there would be corroboration rather than a re-post of the same screenshot.

The row is a receipt, not a release. What it receipts is more interesting than the headline: a model that is strong at low effort, peaks at high, and then spends five times the budget to get worse. If that curve belongs to Anthropic's next flagship, the company has a reasoning-scaling problem to solve before it has a launch. If it belongs to an early checkpoint, the number will move before the name does. Either way, 46.8 is not the figure to watch. The figure to watch is whether the curve ever becomes monotonic.

FAQ

Has Obsidian 3 been released?

No. No model called Obsidian 3 appears in Anthropic's documentation or in Cognition's published FrontierCode 2.0 table, and we found no API catalogue listing it. The only evidence is a hidden row reported in X posts on 27 and 28 October and a forum thread that reproduced five effort scores. Anthropic has not commented, and the public data file we downloaded on 29 October does not contain the row.

Is Obsidian 3 Claude Fable 5.5 or Claude Opus 5.6?

Unknown. Both are guesses from the claude-code harness label, not from any identifier. Fable 5.5 fits the empty slot above Claude Fable 5.1 in Anthropic's lineup; Opus 5.6 fits the theory that Anthropic ships low-gain mid-cycle refreshes. Neither name appears in Anthropic's documentation, and the cost profile fits a premium Fable tier more than a value Opus one.

Why is Obsidian 3 missing from FrontierCode 2.0 now?

We cannot confirm why. The file linked in the original post was last modified at 17:44 UTC on 28 October, and the copy we downloaded on 29 October lists 24 models with no Obsidian 3 entry. It may have been removed after the posts circulated, or the public build may never have contained it. The X posts and the forum archive are the only record we could find, and neither is first-party.

Does the 3.6× cost claim hold up?

The arithmetic does, against Cognition's published Claude Opus 5.5 costs: $8.62 is 3.56× the $2.42 high-effort run and about 3.67× the $12.82 max run when compared with the roughly $47 max figure. What cannot be checked is the Obsidian 3 side of the ledger, because its cost figures exist only in X posts, not in any downloadable file.