Summary

A same-day reaction video, filmed from a hotel room at Meta Connect, covering two frontier model launches hours apart: Anthropic's Claude Opus 5.5 in the morning and OpenAI's GPT-6 Sol and Luna about two hours later. Wolfe's structure is benchmarks first, then third-party demos, then a verdict — and he states up front that he weights benchmarks less than he used to and trusts what he can see people building. The verdict is lopsided: Opus 5.5 beats the tier above it in Anthropic's own lineup for substantially less money, while GPT-6 Sol is cheaper and faster than its predecessor but does not reach OpenAI's own state-of-the-art. The video is worth logging partly because it contradicts itself and says so — Wolfe's own automated leaderboard ranks the two models in the opposite order to his verdict, and he tells viewers to disregard it.

• Anthropic's published claim as Wolfe reads it from their site: Opus 5.5 "now performs at the level of Claude 5.1 on most work and costs 40% less to run than Opus 5". Note that the vendor's claim is parity with the higher tier, while Wolfe's own reading of the benchmark table is stronger — that Opus 5.5 beats Fable 5.1 and Opus 5 across the board. The gap between vendor claim and creator interpretation should not be collapsed.
• Benchmark figures claimed on Anthropic's chart, all reported not verified: agentic coding 66.4% (state of the art), agentic coding on frontier code 54.4% (state of the art), agentic coding on Cursor Bench 57.8% (state of the art), GDPval for knowledge work 1846 (state of the art), plus new state of the art on Humanity's Last Exam, computer use and visual chart recognition. Automation Bench and terminal-bench science are both still credited to GPT-6 Astra, so the sweep is not total and Wolfe says so.
• Pricing claimed: Opus 5.5 at $4 per million input tokens, down from $5, and $20 per million output, down from $25, against Opus 5. Against the tier above — Fable 5.1 at a claimed $10 input and $50 output — that is roughly 60% less per token for capability Anthropic describes as comparable.
• Artificial Analysis aggregate score claimed: Opus 5.5 at 58, taking the lead; the previous best Fable 5.1 at 53, tied with GPT-6 Astra also at 53. GPT-6 Sol comes in around 48 on the same index, tied with Muse Spark, with GPT-6 Luna level with the previous generation's Luna.
• The per-token price cut does not translate into a proportional drop in cost per task, because Opus 5.5 uses far more output tokens per task: a claimed 119,000 against Fable 5.1's 78,000, with GPT-6 Sol at about 31,000. The dollar-per-task figures in this transcript are mangled by the auto-captions — they appear as "$763", "$5.98" and "a$16" — so the token counts and the direction of travel are reportable here and the per-task dollar amounts are not.
• Demos, which Wolfe treats as the real evidence: JavaScript animations from Drew and from Kevin, a Game Boy emulator built and then played ("the Game Boy test"), a game from Edwin, and a Blender claymation-style render from Alex Albert in which the model rendered the video and checked the frames before sending it. Wolfe discloses that Edwin and Alex Albert work at Anthropic, which is the right caveat to keep — a meaningful share of the impressive demos came from the vendor's own staff. Independent examples he cites are a first-person doodle shooter, a Snake variant, and from Matt Berman's Forward Future team a Dark Souls-style game, a flight simulator and a Mario Maker clone. His own assessment is explicitly impressionistic: he says a year ago models were not making animations at all, and that if someone had shown him the Dark Souls clip a year ago he would have assumed he was being trolled.
• OpenAI's GPT-6 Sol and Luna, launched roughly two hours after Opus 5.5, are described as good but not state of the art, with a blog post light on benchmarks and the few shown cherry-picked. Claimed pricing halved against 5.6 Sol: $4 to $2 per million input, $20 to $10 per million output; Luna from 20 cents to 10 cents input and from $1.20 to 50 cents output. On the benchmarks shown, Sol scores below Astra on Automation Bench, on an agents' exam benchmark and on the coding benchmark the captions render as "deep suite", though at lower cost per task in each case.
• Availability: Opus 5.5 available essentially everywhere immediately, including in Claude. GPT-6 Sol and Luna available in ChatGPT, work and Codex from launch day on paid plans (plus, pro, business, enterprise, .edu), with free and lower-tier users getting Luna in the desktop app.
• Wolfe observes that OpenAI appears not to have seeded early access for these models — he could find no third-party demos at all — whereas Fable had visibly had early access partners who published demos at launch. He is careful to say he had early access to neither. This is a distribution observation, not a quality one, and it partly explains the asymmetry in what he was able to show.
• The video contradicts itself and flags it. Wolfe ran both models through his own LLM-as-judge leaderboard and it ranked GPT-6 Sol first, GPT-6 Sol Pro just below, and Claude Opus 5.5 eighth — the opposite of the verdict he delivers. His response is to tell viewers to take the leaderboard with a grain of salt, that he personally thinks Astra still looks better than GPT-6 Sol, and that he pays attention to his eyes rather than the ranking. The disagreement between his automated measure and his stated conclusion is left unresolved in his favour.
• Wolfe's closing summary: Anthropic shipped a model cheaper, faster and better than its own previous state of the art, which implies their next Fable will be better still; OpenAI shipped what he calls a fairly marginal update — cheaper and faster than the previous Sol generation, but not as good as Astra — likely to matter mainly to developers wanting a cheaper GPT model in the API.

Why it matters

Two frontier launches on the same day, and the useful content is the shape of the comparison rather than any single score. Three things are worth keeping. First, the pricing structure has inverted in a way worth watching: a second-tier model is claimed to match or beat the vendor's own top tier at roughly 60% less per token, which makes the tier label a poorer guide to capability than it was, and Wolfe's token-per-task figures show why the headline price cut does not reach the bill — the cheaper model thinks for longer. Second, this entry is a clean case study in how to read a launch-day review: the benchmark numbers are vendor-published, several of the most persuasive demos came from the vendor's own employees (which the creator does disclose), and the absence of demos for the OpenAI models reflects early-access distribution rather than measured quality. Third, and most useful for source discipline, the creator's own automated leaderboard ranked the models in the opposite order to his verdict and he set it aside on taste, explicitly. That contradiction belongs in the record: it is a worked example of a reviewer preferring impression over his own measure, and any downstream claim that Opus 5.5 "won" this day should carry it. Everything here is reported on launch day from vendor materials and social demos, with no independent testing.

0t-eWrGFZyA-transcript.txt