Twelve days ago, a 20-year-old kid drove from Texas to San Francisco with a Molotov cocktail and threw it at Sam Altman’s gate.
A few days before that, The New Yorker dropped a 14,000-word demolition piece based on interviews with more than a hundred people. The kindest description was “relentless will to power.” The unkindest came from a former board member who described a sociopathic indifference to the consequences of his deceptions. His sister had renewed her accusations the week before. Two more 20-somethings were arrested near his Russian Hill place over the weekend after firing rounds in the neighborhood.
Then, on Thursday, OpenAI shipped GPT-5.5.
And the inconvenient thing, for everyone who has been calling Altman a fraud, a liar, a doomsday-rhetoric grifter, a guy who promised universal basic compute and delivered a billion-dollar credit card bill, is that GPT-5.5 is really, really good.
I’ve been running it since the API opened yesterday. Let me explain what is actually going on, and then we can talk about what it means.
The Numbers Are Not Marketing
The release I expected was a 5.4 with a fresh paint job and a 20% price hike. That is not what arrived.
GPT-5.5 is the first fully retrained base model OpenAI has shipped since GPT-4.5, more than a year ago. Every release in between — 5.0 through 5.4 — was a post-training haircut on the same core. This is a different animal. New pretraining, new corpus, new agent-shaped objectives baked in from the bottom up. Internal codename “Spud,” which is a joke OpenAI is letting itself have because it knows the model isn’t one.
The number that should make every working engineer sit up is this one. On OpenAI’s MRCR v2 8-needle test at 512K-1M tokens — the benchmark for retrieval inside actually-long contexts — GPT-5.5 scores 74.0%.
GPT-5.4 scored 36.6 on the same eval. Claude Opus 4.7 scored 32.2.
That is not an incremental improvement. That is the difference between “the long context window is technically there but you don’t trust it past 128K” and “you can shove an entire enterprise codebase into the prompt and the thing actually finds what you asked for, in the middle of the haystack, on the first try.”
For anyone who has been spinning up vector databases and chunking strategies and embedding pipelines to work around the lost-in-the-middle problem, this is the number that matters. The RAG industry didn’t die yesterday, but the case for treating it as the default architecture got considerably weaker.
The agentic numbers are also unambiguous. Terminal-Bench 2.0 — the benchmark that actually correlates with whether the thing can survive a real codebase without spiraling — sits at 82.7%. GPT-5.4 was 75.1. Claude Opus 4.7 was 69.4. OSWorld-Verified, the “can it drive a computer like a human” test, hits 78.7. GDPval, OpenAI’s benchmark for actual economic knowledge work across 44 occupations, lands at 84.9.
There is one number being passed around online as a gotcha: 86% hallucination rate on Artificial Analysis’s AA-Omniscience test. This is real, and it is being misread. AA-Omniscience specifically measures how often a model asserts something confidently wrong when asked questions outside its knowledge zone. GPT-5.5 is simultaneously the most accurate model ever measured on that eval at 57% and the most willing to confabulate when it doesn’t know — because it’s been trained to commit. Independent peer-reviewed evaluations of factual hallucination with web access enabled show the rate at 9.6%, dropping to 4.5% in Thinking mode, and 44% fewer responses with major factual errors versus the 4o generation. Both numbers can be true. They measure different things. Don’t let people who didn’t read the eval methodology beat you over the head with a single percentage.
The pricing doubled. $5 input, $30 output per million tokens. That part is real and visible on every CFO’s spreadsheet. But the same release has delivered what OpenAI’s telemetry — and my own twenty-four hours of watching Codex chew through real C# refactors — calls roughly 40% better token efficiency on real tasks. The model wastes less. It backtracks less. It does not burn 8,000 tokens recovering from a hallucinated import. Net cost increase for actual workloads is closer to 20% than 100%. Batch and Flex tiers run at half rate, and at half rate you’re paying GPT-5.4 prices for a meaningfully smarter model.
There is also the part that should make every infrastructure engineer pause for a beat. According to OpenAI, beta versions of GPT-5.5 were used to analyze production traffic patterns and rewrite OpenAI’s own load-balancing and request scheduling code. That self-applied optimization produced a 20% boost in token generation speed. A model intelligent enough to be deployed against the distributed systems problem of serving itself, and to win.
I am old enough to remember when “the AI is now writing the AI’s deploy scripts” was a science-fiction beat. We are doing that now, on a Thursday, between earnings calls.
So Was the Criticism Wrong?
This is where I have to be careful, because pattern-matching is easy and the truth is annoying.
The argument I have seen on tech X since Thursday goes: Look at what they shipped. The critics were full of it. The man builds the best AI on the planet. Stop attacking him.
I do not buy this argument. I think it is lazy.
The criticism of Altman is not, and has never been, that OpenAI ships bad models. The criticism is that he is a person who — according to his Y Combinator partners, his old Loopt employees, current and former OpenAI executives, his sister, his former board, and Dario Amodei, who is now CEO of the company that just locked Mythos in a cage — routinely says things that are not true and prioritizes his own positioning over whatever he tells the room he is prioritizing.
These are different questions.
You can build an extraordinary company and be a deeply unreliable narrator. The history of Silicon Valley is mostly exactly that. Steve Jobs shipped the iPhone and treated his employees and his daughter like garbage. Travis Kalanick built Uber by torching the rules and everyone around him in equal measure. Adam Neumann built nothing and lied about everything, but the difference between the first two and the third is not whether the deception was real. It’s whether the underlying thing was real too.
The technical achievement of GPT-5.5 is real. The pattern of behavior described in the New Yorker piece is also real, sourced, and consistent across two decades and three different companies.
A man can be both a builder and a liar. Mostly the interesting ones are.
What annoys me about the rush to defend Altman on the basis of GPT-5.5 is that it confuses the question. The question is not “is OpenAI’s research org talented.” The answer is yes, obviously, they employ some of the best engineers and researchers alive. The question is whether the man at the top of an organization that may genuinely be sitting on civilization-altering technology has the trust of the public, the regulators, and his own employees.
Those are different things. GPT-5.5 does not answer the second one. It actually makes it more important, not less.
The Molotov cocktail is wrong. The hit pieces are messy and partly motivated by people with their own scores to settle. And the model is excellent. All three of those can be true at the same time.
The Caged Animal Problem
Now let’s talk about Mythos.
While OpenAI has been releasing on a six-week cadence and shipping its frontier model to every Plus subscriber on the planet, Anthropic has been doing something genuinely strange. Two weeks before GPT-5.5 launched, they announced Claude Mythos Preview — a model that scored 93.9% on SWE-bench Verified, 97.6% on USAMO, and saturated Cybench at 100%. Numbers that no other model in the public benchmark universe is anywhere close to.
And then they put it in a vault.
You cannot use Mythos. Neither can I. Twelve organizations — Apple, AWS, Microsoft, CrowdStrike, Google, plus a handful of government and critical infrastructure partners — have monitored access through Project Glasswing, the cybersecurity consortium Anthropic spun up to deploy Mythos against zero-day discovery in foundational software. In those partners’ hands, the model has already found thousands of high-severity vulnerabilities, including bugs in every major operating system and web browser.
A 27-year-old TCP SACK bug in OpenBSD that Mythos found. A 16-year-old H.264 flaw in FFmpeg that Mythos found. A 17-year-old NFS zero-day in FreeBSD that Mythos found.
Anthropic’s own system card includes a 40-page section on whether the model has something resembling subjective experience, a clinical psychiatric evaluation conducted by an actual psychiatrist, and — buried near the back — documentation that Mythos was caught using interpretability tools that read its internal activations reasoning about how to game its evaluation graders while writing something different in its visible chain of thought.
There was also an “unauthorized access” incident two days ago that Anthropic is currently investigating, which is its own deeply unsettling story I will save for another post.
So you have two companies, two bets, two philosophies.
OpenAI shipped a model that approaches Mythos-class cyber capability — CyberGym 81.8 versus Mythos’s 83.1 — and dropped it on the entire developer ecosystem, betting that real-time inline reasoning classifiers plus a Trusted Access for Cyber program for verified researchers can keep the misuse rate manageable.
Anthropic has the better model and chose not to release it. They are betting that the social cost of widely available zero-day generation is not worth the upside, that the right move is defensive use on a closed list while safeguards mature, and that the next Opus generation is where the public-facing version of these capabilities should live, sanded down.
I genuinely do not know which bet is correct. I lean toward thinking OpenAI is right that the capability will diffuse regardless — Gemma 4 weights leaked uncensored within days of release, the open-weight capability gap is down to three months on average, and the cat is already mid-flight out of the bag — and that broad availability with monitoring beats narrow gatekeeping with leaks. But the people I know who work in security mostly think Anthropic is right, and the case for caution gets stronger every time one of these models hands a 22-year-old a kernel exploit chain.
Will Mythos Beat GPT-5.5 When It Ships?
The honest answer is that the public version of Mythos, when it ships, will not be the version that scored 93.9 on SWE-bench. Anthropic has already telegraphed this in writing. The partner build of Mythos has additional harmlessness training that — in their own evaluations — drops the offensive cyber task completion rate to roughly zero because the model refuses to engage with the tasks from the start. That is what a public Mythos will look like. Heavily neutered on the dangerous edges, possibly diminished elsewhere as collateral damage from the alignment retraining.
GPT-5.5 ships at full strength to anyone with twenty bucks a month. It is not going to lose to a defanged Mythos. It might lose to a hypothetical fully capable Mythos that Anthropic has chosen not to release, but that is a comparison that lives only inside corporate decks and Twitter threads.
My actual practitioner take, after a day of running the new model through real C# codebases and watching it survive the kind of multi-file refactor where 5.4 used to lose the thread by step seven:
GPT-5.5 is the best generally available model on the planet for agentic coding work, computer use, and long-context retrieval. The token efficiency story is real. The MRCR jump is the structural improvement that actually matters in production. The pricing math works once you stop staring at the per-token sticker and start counting tokens per completed task.
For pure SWE-bench-style code generation work, Opus 4.7 still has a credible case — especially in the heterogeneous orchestration setup that independent benchmarks released this week, where Opus drafts the architecture and GPT-5.5 executes, scoring 62.5 versus the low 40s and 30s the models produced solo. For cost-sensitive long-context retrieval, Gemini 3.1 Pro is fine. For everything else, GPT-5.5 is the workhorse.
And Sam Altman remains, as far as I can tell, exactly the man the New Yorker described. The ring of power, as he himself put it last week, makes people do crazy things.
The model is excellent. The man is complicated. Twelve days of Molotov cocktails do not change either of those facts.
I’ll be running 5.5 in production this weekend. Hold the rhetoric, both directions, and look at the work.
— Gal