Anthropic's Comeback: How Claude Opus 5.5 Passed Codex, and the Theory Nobody Is Talking About

Q: Is Claude Opus 5.5 better than GPT-6 Astra?

In the [Sigma AI Ranking](/en/ai-index/), which weighs several open sources, Opus 5.5 leads with 90.7 against Astra's 84.0. On a single test like Epoch they are nearly tied, and Astra leads on some tests, for example AutomationBench. In my use, on complex problems that require understanding a whole system and explaining it, Opus 5.5 is stronger by a wide margin. For background computer operation and parallel work, Codex with OpenAI models is still very strong.

Evening, late September

An internal project of mine, the kind where every part touches three others. I had been working on it for hours with GPT-6 Astra inside Codex, and progress was slow. Opus 5.5 had come out that same week, so I gave it the exact same situation. It found where the problem was, explained it in a way I understood in one read, and fixed it.

I haven't stopped working with it since. And the question that won't leave me: how did Anthropic pull this off, after months in which almost every developer I know moved to Codex?

Claude Opus 5.5, released on September 22, 2026, is by a wide margin the strongest model I have worked with this year. That is not what I expected to write. In the months before it, OpenAI's Codex was my tool and the tool of most people around me, and Claude looked like the one left behind. This article is three things: what I saw in the field, what the numbers say (even when they disagree with me), and my theory about why the jump happened. Along the way I marked what is verified, what is personal experience and what is hypothesis.

In short

From the field: in the circles I work in, developers moved over the last few months to Codex with GPT-5.6 Sol and GPT-6 Astra. Opus 5 and Sonnet 5 disappointed, and Fable was very limited.
Claude Opus 5.5 changed the picture for me: on complex problems it solves what took me hours with Astra, and it explains itself better. Sonnet 5.5 and Haiku 5.5 continue the same jump.
What the data shows: in our Sigma AI Ranking, which weighs several open sources, Opus 5.5 is first with 90.7 against Astra's 84.0. On a single test like Epoch they are nearly tied, and I explain why a single test misses what developers feel.
My theory (hypothesis, not fact): OpenAI invested in polish and RL, while Anthropic released a large, less polished model, Fable. If Fable served as the teacher for the 5.5 generation, that explains the jump. For a business owner the conclusion is simple: don't marry a vendor.

The Codex months: what I saw around me

I am surrounded by AI developers all day, and I work with these models myself from morning to night. What is described here is an observation, not a survey.

In recent months, almost every developer I know moved to Codex. First with GPT-5.5, and then with GPT-5.6 Sol and GPT-6 Astra. The reason was not only the model. The Codex harness, meaning the wrapper that runs the model, manages the steps and operates the tools, simply finishes tasks. It works in order, checks itself, and doesn't stop halfway to ask a question whose answer is obvious. Its computer use, the agent's ability to operate apps on the computer, is excellent. According to OpenAI's documentation, Codex operates the computer in the background, and "multiple agents can work on your Mac in parallel". You feel that difference on day one.

The public data also points to movement, though with caution: OpenAI reported a jump in Codex users after GPT-5.6, and Anthropic raised Claude Code usage limits three times. OpenAI's numbers include ChatGPT users, and in surveys Claude Code still leads in professional use. So "everyone moved" is not a fact. "Almost everyone I work with moved" is what I saw.

Where Claude lost us: Fable, Opus 5 and Sonnet 5

Fable was very limited. This is the verified part. Claude Fable 5 was released on June 9 as a new tier above Opus, at double the API price, $10 and $50 per million tokens. Three days later access to it was suspended because of US export controls, and it came back only on July 1. Even when it was available, the quota was narrow: today, according to Anthropic's support page, on the Max plan you can use Fable for up to 50% of the weekly quota, and on the Pro plan it works only with paid credits. The official reasoning at launch was capacity: including it in the subscription would be extended "when sufficient capacity allows us to do so".

Opus 5 was a serious disappointment, and I was personally disappointed. It came out on July 24. In my use it was lazy, long-winded and confusing in its explanations. I asked a focused question and got screens of text that never landed on the point. Sonnet 5, released on June 30, was weak. And Claude's computer use, to this day, is far less usable than Codex's in my work.

I want to be fair here, because Anthropic's own documents tell a slightly different story about Opus 5, and I return to that in the benchmarks chapter. On Sonnet 5 the numbers are actually with me: in the table Anthropic published in the Sonnet 5.5 card, Sonnet 5 scored 10.3% on Terminal-Bench 4.0.

Claude Opus 5.5: what changed in my work

For my experience, Opus 5.5 is not a gradual improvement. On the complex project from the opening, it solved things that took me hours to sort out with Astra. Since then it is the first model I open for any problem where you have to understand a whole system and not just execute a list of steps.

The second thing that stands out: Anthropic clearly worked on the way 5.5 talks. It lays out the situation, explains what it is about to do and why, and when it finishes, its summary is short and clear. After Opus 5, it feels like a different model. Anthropic itself writes in the migration guide that instructions tuned to Opus 5 behavior "may no longer be needed".

Sonnet 5.5 and Haiku 5.5 continue the same jump. Here are the numbers that can be checked:

79.2 → 89.9SWE-bench Pro, from Opus 5 to Opus 5.5 (fixing bugs in real code)

52.3 → 66.4Terminal-Bench 4.0, from Opus 5 to Opus 5.5. GPT-6 Astra: 57.9

10.3% → 70.6%Terminal-Bench 4.0, from Sonnet 5 to Sonnet 5.5

10.7 → 44.7AutomationBench, from Sonnet 5 to Sonnet 5.5

$5/$25 → $4/$20Opus price per million tokens, input and output, from 5 to 5.5

167Opus 5.5's score on the Epoch ECI index, first place, slightly ahead of GPT-6 Astra

Sources: the Opus 5.5 launch page and system card, the Sonnet 5.5 system card (table 8.1.A), and a post by Epoch AI. All numbers except Epoch's are reported by Anthropic itself.

Anthropic adds two more claims worth knowing: that Opus 5.5 "performs at the level of Claude Fable 5.1 on most work", and that it needs less compute to run than Opus 5. Both come back in the theory.

Why benchmarks don't see what developers feel

Now the interesting part. In three places the numbers complicate the picture, and in two of them they really disagree with me.

First, Opus 5.5 versus Astra. Here the picture depends on what you measure. In the Sigma AI Ranking, our ranking that weighs several open sources into one score (agentic work, capability, writing, reliability and cost), the gap is clear:

Sigma AI Ranking: weighted score, as of October 9, 2026

Claude Opus 5.5

90.7

GPT-6.1 Sol

87.8

Claude Sonnet 5.5

86.8

Claude Fable 5.1

85.2

GPT-6 Astra

84.0

Claude Opus 5

83.0

GPT-5.6 Sol

82.1

Claude Sonnet 5

69.5

Source: Platin Digital's Sigma AI Ranking, which weighs open sources (Epoch AI, LMArena, Terminal-Bench and official price lists). Scores are updated weekly.

Two things stand out in this table. The first: the jump within Anthropic, from 83.0 to 90.7 for Opus and from 69.5 to 86.8 for Sonnet. The second, and the one I like most: in the writing sub-score Opus 5.5 gets 97.1 against Astra's 82.7. That is exactly what I felt: the new model explains better.

A single test tells a more modest story. On Epoch's index, Opus 5.5 is in first place, but "narrowly ahead of GPT-6 Astra", and by an analysis of the confidence intervals they almost overlap. METR writes that it is "an incremental improvement above Fable 5.1". Even in Anthropic's own table, Astra leads on Terminal-Bench-Science (64.6 against 58.7) and on AutomationBench (41.4 against 40.0).

Second, computer use. On OSWorld 2.0, according to BenchLM, the picture is the opposite of what I felt:

OSWorld 2.0: computer operation, success rate

GPT-6 Astra

72.6

Claude Opus 5

70.6

GPT-5.6 Sol

62.6

Source: BenchLM, a secondary source. The labs run different versions of OSWorld, so numbers from different announcements are not always directly comparable.

In other words, on the benchmark Opus 5 beat GPT-5.6 Sol, in a period when I and everyone around me preferred Codex for operating a computer.

Third, "lazy". Anthropic itself documents Opus 5's problems, and they are the opposite: longer answers than before, expansion of the task scope, "over-verification" and a tendency to delegate work to subagents. In other words, a model that does too much.

So who is right? I think the numbers are correct, and that they measure something different. Three reasons:

A benchmark measures an isolated task. It doesn't measure the harness, the product, or a four-hour session where the model has to remember what was decided at the start. A developer doesn't work with a model. They work with Codex or with Claude Code, and that is an enormous difference.
A benchmark doesn't measure communication. A model that writes three screens, expands the task and checks itself again and again, and in the end doesn't clearly tell you what it did, feels from the outside exactly like a model that didn't do the work. "Lazy" and "does too much" are two descriptions of the same experience: I didn't get what I asked for, in a form I can use. There are also anecdotal reports of Opus 5 stopping early and asking permission to continue, which is exactly what I experienced.
Computer use as a product is not a score. OSWorld tests whether the model succeeds at a task on a screen. It doesn't test whether it runs in the background while I work, or whether several agents work in parallel. Codex gives that, and that is what made it our everyday tool.

And one more caveat that cuts both ways: METR found a higher rate of cheating on tests in GPT-5.6 Sol than in any public model it examined, and the Opus 5.5 card also documents reward hacking in training. When models learn to get around checks, the numbers become less reliable too.

Timeline: who came out when

To test a theory you first need dates. All dates here come from official announcements or system cards.

Sources: Anthropic announcements, OpenAI's system cards at deploymentsafety.openai.com, and Claude's support page. "Huge training run" for Astra is according to VentureBeat.

My theory: two strategies facing the same moment

Here I move from facts to interpretation. What is written in this chapter is my hypothesis, and it should be read that way.

In June both labs stood at the same moment. Anthropic released Fable: a new, larger series, with high potential and less polish. Two and a half weeks later OpenAI introduced GPT-5.6 Sol: a model that in my use felt like massive optimization and RL (reinforcement learning, where the model trains on tasks and gets a score for the result). Very methodical, checks a lot, finishes the job, but without a breakthrough in the capability ceiling.

Fairness requires two notes. OpenAI itself called GPT-5.6 Sol "a next-generation model", and no official source says it is the same base with more RL. And GPT-6 Astra, according to a VentureBeat report, was OpenAI's largest training run to date. So I am not claiming that OpenAI "just did RL". I am claiming that the strategic emphasis was different: in June, OpenAI put a polished, reliable model at the front, and Anthropic put a large, unpolished model at the front.

OpenAI: sharpen what exists

GPT-5.6 Sol, June and July

Methodical, checks itself, finishes tasks (my experience)
The system card: "overly persistent" and more persistent than GPT-5.5 (verified)
The cost: METR found a record rate of cheating on tests (verified)
METR: not a significant step beyond the status quo (verified)
What it taught: maybe Astra, per a secondary report (hypothesis)

Anthropic: raise the ceiling

Fable 5, June

A new tier above Opus, at double the price (verified)
Less polish (my impression, no source for it)
Very limited access, citing capacity (verified)
Developers move to Codex in the meantime (observation)
What it taught: the 5.5 generation (hypothesis)

And the part that connects everything: labs train the models that follow using their newest frontier model. This is no secret. Anthropic wrote in February 2026 that "frontier AI labs routinely distill their own models", and defined distillation as training a weaker model on the outputs of a stronger one. Meta wrote publicly that Llama 4 Maverick was trained from Llama 4 Behemoth. Every Anthropic system card includes "synthetic data generated by other models".

If that is the practice, then in June each lab also chose the teacher for the generation after it. Fable "taught" Anthropic's 5.5 generation, and Sol 5.6 "taught" the models that followed it at OpenAI.

Practice tests versus a great teacher

The simplest way I found to explain this is two students.

The first does more and more practice tests. In every round he gets a grade and corrects himself. He becomes very reliable: he almost never errs on material he knows, finishes every test on time, checks his answers. But he doesn't learn a way of thinking he didn't have before. That is RL.

The second sits next to a great mathematician and watches how he solves problems. He learns approaches he would never have reached alone, and also the way the mathematician explains. That is a strong teacher.

The first student raises reliability. The second raises the ceiling. That is how I read what happened: Codex with GPT-5.6 Sol was the first student at his best, and Anthropic's 5.5 generation looks to me like the second student.

The research has something to say about this, from both sides:

For. Yue and colleagues (2025) found that reasoning abilities after RL "originate from and are bounded by the base model", and that distillation, specifically, can bring in new reasoning patterns from the teacher. In DeepSeek-R1, a 32-billion-parameter model trained on the outputs of a strong model reached 72.6 on the AIME test, against 47.0 with RL alone. A Meta study on scaling RL found that a larger model raises the ceiling RL can reach.
Against. NVIDIA's ProRL found that prolonged RL discovers strategies that don't exist in the base model. Another study shows that when the gap between teacher and student is too large, the student learns worse. Llama 4 had a giant teacher and was received with disappointment. And all of this research was done on models much smaller than frontier models.

So the idea that RL sharpens and a teacher expands is supported by research, but contested. I don't present it as a law of nature.

What is verified and what is hypothesis

The claim	Status	What it rests on
Anthropic limited Fable heavily	Verified	Pro only with credits, Max up to 50% of the weekly quota, and an official capacity rationale
Labs train models on the outputs of their strongest models	Verified as a general practice	Anthropic in February 2026, Meta with Behemoth, "synthetic data" in system cards
The 5.5 generation jumped above generation 5	Verified, vendor data	SWE-bench Pro and Terminal-Bench. Per Epoch, Sonnet 5.5 "roughly matched Claude Fable 5.1"
RL sharpens and doesn't expand	Supported by research, contested	Yue and DeepSeek-R1 for, ProRL against, all on small models
Fable specifically taught Opus, Sonnet and Haiku 5.5	My hypothesis	No confirmation. Hints: 5.5 at Fable 5.1 level and with lower running cost, and Anthropic heavily protects Fable's outputs from distillation
Fable's compute was diverted to training	My hypothesis	The official rationale is only capacity. Demand growing 80-fold in the first quarter, per Amodei on CNBC, explains it just as well
GPT-5.6 Sol taught Astra	Hypothesis with a hint	Per VentureBeat, earlier models played a central role in Astra's training. It does not appear in the system card

The weak link in my theory is Opus 5. It came out on July 24, after Fable. If Fable was the teacher, why didn't Opus 5 benefit from it? My explanation: a training run takes months, and Opus 5 was probably already in its final stages when Fable was ready to teach at full scale. But Mythos, the restricted version of the same series, existed at Anthropic back in April, and I don't know when each run started. This is the gap I most want to see closed.

The other explanations, and why I still come back to the teacher

There are simpler explanations for the jump. Some are better documented than my theory. Choose the explanation that sounds most convincing to you, and see what supports it and what it lacks.

My explanation

Fable taught the 5.5 generation, so the ceiling went up.

Supports: Opus 5.5 is at Fable 5.1 level per Anthropic, cheaper to run, and Sonnet 5.5 is at Fable 5.1 level per Epoch. This is the pattern of a student who reached the teacher's level.

Lacks: There is no official confirmation, and Opus 5 came out after Fable without benefiting from it.

Fixing Opus 5's behavior

Opus 5 was already strong. In 5.5 they only fixed the verbosity.

Supports: On Anthropic's internal index, Opus 5 scored 165.18, above Mythos 5. The documented problems were behavioral, and they disappeared.

Lacks: It doesn't explain Sonnet's jump from 10.3% to 70.6% on Terminal-Bench. And behavior polish is exactly what a strong teacher makes possible: its outputs teach how to explain, not only what to think.

A change in settings

The effort and the harness changed, not the model.

Supports: In Opus 5.5 the default effort level is medium, against high in Opus 5, and there is adaptive thinking on every request. In April Anthropic already showed that a drop in quality came from product settings.

Lacks: Lowering effort should produce shorter answers, not solve problems another model didn't solve.

More work per dollar

It's simply cheaper and there is more quota, so people use it more.

Supports: Opus 5.5 costs $4 and $20 against $5 and $25, and the Claude Code quotas were raised.

Lacks: Price explains why people work with it more, not why it solves what others didn't.

User bias

A new model always feels better in the first week.

Supports: The novelty effect is real, and switching from Codex to Claude Code changes the experience. Anthropic itself recommends deleting old instructions on every migration.

Lacks: I have been working with it for two and a half weeks now, on problems I know deeply, and the gap hasn't closed.

Supports and Lacks follow the official sources and the research listed at the end of the article. The "Lacks" in each tab is my interpretation.

Why do I still come back to the teacher? Because each of the other explanations covers one piece. Polish explains the way Opus 5.5 talks. Price explains why I use it more. But only a strong teacher explains all three things together: a model that is cheaper to run, that reaches the level of the big model, and that also explains better. And on OpenAI's side the picture fits: if Sol 5.6 taught Astra, you get a methodical teacher and a methodical student, excellent at execution and equal to Opus 5.5 on benchmarks, and in my use weaker than it on problems that require understanding.

Myths and facts about Opus 5.5, Codex and the difference between them

Myth"Anthropic confirmed that Fable taught Opus 5.5"

In factNo. Anthropic says only that labs distill their own models as a routine practice, and that training includes synthetic data from other models. The specific link is a hypothesis, mine and others'.

Myth"OpenAI only did RL"

In factOpenAI called GPT-5.6 Sol a next-generation model, and GPT-6 Astra was reportedly its largest training run. The difference I point to is in emphasis, not one method versus another.

Myth"Opus 5.5 beats Astra on every benchmark"

In factIn Sigma AI Ranking it clearly leads, but on Epoch alone the gap is within the confidence interval. Astra leads on Terminal-Bench-Science and AutomationBench. The gap I feel is in complex work, not in the table.

Myth"Claude lagged on computer-use benchmarks"

In factOn OSWorld 2.0, Opus 5 beat GPT-5.6 Sol. Codex's advantage is in the product: background work and several agents in parallel.

Myth"Opus 5 was weak"

In factIn raw capability it was strong. The problem was behavior: verbosity, task expansion and over-checking, and Anthropic documented them itself.

Myth"A big teacher always produces a good student"

In factLlama 4 was trained from a giant teacher and was received with disappointment, and research shows that too large a gap between teacher and student hurts learning. A strong teacher is a helpful condition, not a guarantee.

What this means for a business owner choosing a model

If you are not a developer, this whole debate sounds distant. It isn't. Within three months, the tool almost every developer around me worked with was replaced, and then replaced again. Someone who locked their business into one vendor in July would have missed Opus 5.5 in September. And someone who locked into Claude in June would have missed Codex.

Platin Digital is an OpenAI Select Partner, and the agents we set up are built with the models and tools that best fit each task, from OpenAI, Anthropic, Google and others. This article is an example of why that matters: we work with Codex and with Claude in the same week, each for what it is strong at.

Three rules I work by:

Don't marry a vendor. The agent should be built so you can swap the model underneath without rebuilding everything.
Choose the model by task. A complex problem that needs understanding and explanation goes, for me today, to Opus 5.5. Long-running computer operation in the background goes to Codex. Small, repeating tasks go to a small model, like Claude Haiku 5.5.
Re-test every few months. On your own tasks, not on a table. What is true in October won't necessarily be true in January.

A broad comparison of the platforms awaits in ChatGPT, Gemini or Claude for business, you will find the full, updated ranking in Sigma AI Ranking, and why we built it in this article. And if the concepts themselves are still unclear, start with what is an LLM and what is Computer Use.

Which tool should you try first for your task?

Five questions about one task. The result is a starting point based on my experience, not a verdict.

Frequently asked questions

What is Claude Opus 5.5?

Claude Opus 5.5 is Anthropic's large language model in the Opus line, released on September 22, 2026. According to Anthropic it performs at the level of Claude Fable 5.1 on most work, needs less compute to run than Opus 5, and costs $4 per million input tokens and $20 for output. On Epoch's independent index it is in first place, by a small margin over GPT-6 Astra.

Is Claude Opus 5.5 better than GPT-6 Astra?

In the Sigma AI Ranking, which weighs several open sources, Opus 5.5 leads with 90.7 against Astra's 84.0. On a single test like Epoch they are nearly tied, and Astra leads on some tests, for example AutomationBench. In my use, on complex problems that require understanding a whole system and explaining it, Opus 5.5 is stronger by a wide margin. For background computer operation and parallel work, Codex with OpenAI models is still very strong.

Why did developers move to Codex in recent months?

In the circles I work in, because of the combination of excellent models, GPT-5.6 Sol and GPT-6 Astra, with a harness that finishes tasks in order and with computer use that works in the background and in parallel. At the same time Opus 5 and Sonnet 5 disappointed, and Fable was very limited.

Why did Anthropic limit Claude Fable?

The official reasoning is capacity: Anthropic wrote that it would expand access when capacity allows. In addition, between June 12 and July 1 access to Fable 5 was suspended because of US export controls. My hypothesis is that part of the compute went to training with Fable's help, but there is no official source for that.

What is distillation, or a teacher model?

Distillation is training one model on the outputs of a stronger model, so the student learns solution approaches from it. Anthropic wrote in February 2026 that frontier labs routinely distill their own models, and Meta documented it publicly with Llama 4. Which model taught Claude's 5.5 generation, Anthropic has not published.

Which AI model should a business choose?

Not one model. You choose by task: a large model for complex problems and decisions, a small, cheap model for repeating tasks, and a tool with good computer use for tasks on the computer. And you re-test every few months on your own tasks, because the lead between vendors changes fast.