GPT-5.6 vs Claude 5: What Actually Changed in 2026

I need to open with a correction rather than a hook, because this post used to be called "GPT-5 vs Claude 4" and both halves of that title are now obsolete.

That is not a small edit. When I went to refresh it I found the site was describing GPT-5 and Claude 4 as current across roughly 220 references, while OpenAI had moved to the GPT-5.6 family and Anthropic had shipped the Claude 5 family. Two generations of drift on each side, in a post about the models themselves.

So this is a rewrite, and it is built differently from the version it replaces. Fewer verdicts, more checkable facts, and one section explaining why I am not giving you the benchmark table you probably came here for.

The quick version if you are in a hurry

The single most useful thing I can tell you is that the question changed shape. Both vendors stopped shipping one flagship and started shipping a family of tiers, so "which is better" now depends far more on which tier you buy than on which company you buy from.

A frontier tier from either vendor beats a budget tier from the other. That makes brand loyalty an expensive way to choose.

If you want one sentence each: Anthropic's line runs Haiku 4.5 for cheap fast work, Sonnet 5 for the balance, Opus 5 for complex agentic coding, and Fable 5 above that for the hardest long-horizon reasoning. OpenAI's runs Luna for cost efficiency, Terra for balanced everyday work, and Sol at the frontier.

What each family actually is now

Verified against Anthropic's model reference and OpenAI's own announcement on August 4, 2026.

Anthropic, the Claude 5 family. Claude Opus 5 arrived on July 24, 2026. Claude Sonnet 5 sits below it, Claude Fable 5 above it as the most capable widely released model, aimed at the most demanding reasoning and long-horizon agentic work. Claude Haiku 4.5 remains the fast, cheap option. There is also a restricted-access model, Claude Mythos 5, which you cannot simply sign up for.

Getting there was not one jump. Opus 4.6, then 4.7, then 4.8 all shipped across the first half of 2026 before the 5 family arrived, which is four Opus-tier releases inside roughly six months. I am describing that sequence rather than dating each release, because the precise dates I found for the 4.x releases came from a single aggregator and I have not confirmed them against Anthropic's own announcements. The Opus 5 date I am confident in.

OpenAI, the GPT-5.6 family. Three named tiers rather than one flagship: Sol at the frontier, Terra as the balanced everyday model, Luna as the most cost-efficient. GPT-5.5 existed in between GPT-5 and this, which is worth knowing if you are reading anything written earlier in the year.

The shared pattern is the interesting part. Both vendors independently concluded that one model cannot serve a coding agent, a customer-support bot and a summarisation job at sensible cost, so both split the lineup. Comparison articles have not caught up with that, and still write as though there are two products.

What they cost

API pricing per million tokens, input and output, read off Anthropic's model reference on August 4, 2026 and re-checked on September 1, 2026:

A correction on that Sonnet 5 row. This page used to show $3 and $15 as the standard rate with $2 and $10 as an introductory price expiring on August 31, 2026, and told you to budget with the higher pair after that date. I went back to Anthropic on September 1, the day after it supposedly expired, and the page lists Sonnet 5 at $2 and $10 with the word introductory appearing nowhere on it. Budget with $2 and $10.

Worth knowing where $3 and $15 comes from, because it is a real Anthropic price: it is what Sonnet 4.6 costs. Whether this page originally conflated the two models or Anthropic restructured the tier, I cannot tell from the outside, and I am not going to invent a reason.

On the OpenAI side I can report a dated movement rather than a full table: on July 30, 2026, GPT-5.6 Luna dropped 80% in price and Terra dropped 20%. An 80% cut on the cheap tier is a real signal about where that vendor expects volume to go.

Consumer plans have barely moved. ChatGPT Plus and Claude Pro are both still $20 a month, ChatGPT Pro is $200, Claude Max is $100, and Team plans on both land in the $25 to $30 per-user range. Full breakdowns in my ChatGPT Plus review and Claude AI review.

What matters more than those numbers is what sits behind them. The story of AI pricing this year was flat subscriptions quietly acquiring metered usage allowances, which I documented change by change in what AI pricing did in 2026. A $20 plan is not the same product it was in January.

What a million tokens is actually for

The context number is the specification people quote most and understand least, so it is worth grounding.

A million tokens is roughly 700,000 words, depending on the text. That is a long novel, or a mid-sized codebase, or a few hundred pages of contracts and their appendices. Whether you need it has nothing to do with how sophisticated your work is and everything to do with its shape.

You need it when the task requires the model to hold a whole thing at once and reason across all of it: find every place a function is called before renaming it, check whether clause 14 contradicts anything in the appendices, trace one character's dialogue for consistency across a manuscript. Chunking those into pieces does not merely slow them down, it changes the answer, because the contradiction you are looking for lives between the chunks.

You do not need it for the majority of what people actually do. Drafting, summarising a single document, answering a question, writing a function. Those fit comfortably in a fraction of the window, and paying for capacity you never fill is just paying more.

Two caveats that comparison tables leave out. Large context is not free: you are billed on tokens read, so a 900,000-token conversation costs roughly what 900,000 tokens costs, every turn, and that adds up faster than people expect. And a big window is not the same as good recall across it, which is a genuine engineering problem rather than a marketing number, and one you should test on your own material rather than trust.

Haiku 4.5's 200,000-token window is the odd one out in the Claude lineup, and that is a deliberate tier decision rather than an oversight. The cheap fast model is for tasks where you were never going to send a novel.

How to pick a tier

Once you accept that the choice is a tier rather than a brand, picking gets easier, because tiers map onto properties of the task rather than onto vibes.

Four questions get you there.

How much does the model need to hold at once? If the answer is a whole repository or a whole case file, you want a 1 million token window and the tier question is settled by that alone. If it is one document, you have the full range available.

Is the work agentic or single-shot? A single request that returns an answer behaves predictably, and a cheaper tier usually handles it. An agent that runs for twenty minutes, reads a hundred files, calls six tools and self-corrects is a different workload, which is why both vendors position specific tiers for it. Anthropic points Opus 5 at complex agentic coding and Fable 5 at long-horizon work; OpenAI points Sol at the frontier.

How many times will you call it? This is the one people get wrong. At a hundred calls a month the price difference between tiers is rounding. At ten million calls it is the entire business model, and the gap between $1 and $10 per million input tokens decides whether the product works. If you are building rather than chatting, start from the cheapest tier that passes your evaluation and move up only where it fails.

Is anyone waiting? Frontier tiers now think before answering, sometimes for a long time on hard problems. That is the right trade for an overnight refactor and the wrong one for a chat widget where someone is watching a spinner. If latency is the binding constraint, drop a tier rather than switching vendor.

Notice that none of those questions is "which model is smarter." That is because for most real work the tier decision dominates, and the frontier-versus-frontier question, the one every comparison article is built around, is the one that matters least to the largest number of readers.

Why there is no benchmark table in this post

You came here for a scoreboard. I am not going to give you one, and the reason is worth more than the scoreboard would be.

There are specific head-to-head numbers circulating for Claude Opus 5 against GPT-5.6 Sol. SWE-bench Pro at 79.2% against 64.6%. Frontier-Bench agentic coding at 43.3% against 34.4%. ARC-AGI-3 at 30.2% against 7.8%. They look authoritative, they are quoted to one decimal place, and they appear across a dozen pages.

I could not trace a single one to a published benchmark run. Every page carrying them is an aggregator, and several of them quote each other, which is how a number acquires the appearance of consensus without anyone having measured anything.

Four days ago I wrote a comparison of two coding tools and refused to repeat SWE-Bench figures I could not source. Repeating a different set of unsourced figures today, because they would make this article more satisfying, would mean the earlier refusal was a pose rather than a standard. So: prices, dates, context windows and lineups are verifiable, and this post is built on those. Capability rankings between current frontier tiers are not something I can honestly hand you.

If you want a number you can trust, the one that matters is how each model does on your work. Run the same real task through both. That takes an afternoon and beats every table on the internet, including any I might have written.

Where each family has a structural edge

Some differences are architectural rather than a matter of degree, which makes them safe to state.

Context. Every current Claude model except Haiku 4.5 carries a 1 million token window. If your work involves dropping an entire codebase, a deposition, or a year of documentation into a single conversation, that capacity is the feature, and it is not something prompting works around.

Image generation. Claude reads images and PDFs but does not appear to generate them: Anthropic's capability documentation lists image input with no image output counterpart. If generation is part of your workflow, that is ChatGPT or a specialist tool. My Midjourney vs DALL-E 3 comparison, updated for the August 30, 2026 DALL-E retirement covers the dedicated options and what ChatGPT's image generation runs on now.

Tier granularity. Anthropic ships four openly available tiers plus a restricted one; OpenAI ships three. If you are building something where per-token cost compounds across millions of calls, having a genuinely cheap tier from the same family, with the same API shape, matters more than the frontier model's ceiling.

Ecosystem shape. ChatGPT bundles more into the consumer product: voice, image generation, browsing, code interpreter, custom GPTs. Claude's consumer product is narrower and deeper on text and code. Neither is better in the abstract; they are different products at the same price.

For the everyday version of this comparison rather than the model-generation version, Claude vs ChatGPT is the more practical read, and ChatGPT vs Google Gemini covers the third option most of these articles skip.

Speed and reliability

Both are fast enough that the model is rarely what you are waiting for, and both have had incidents this year. What has genuinely changed is that the frontier tiers now think for longer before answering on hard problems, which reads as slowness if you expect a chatbot and as thoroughness if you expect a colleague.

If latency is the constraint, the answer is a cheaper tier, not a different vendor. Haiku 4.5 and Luna exist for exactly that.

Which one should you get

Pick on your dominant task, not on a leaderboard. And if a single subscription has to cover everything, use the other one's free tier for the gaps, which costs nothing and covers most of them.

My setup

Claude Pro as the daily driver, because most of my day is writing and code and the long context earns its keep on both. ChatGPT free alongside it for image generation and the occasional voice conversation. For coding specifically I work through an editor rather than a chat window, which is a separate decision I worked through in Cursor vs GitHub Copilot, and my walkthrough of using Claude for coding covers the terminal-first approach.

I have changed this setup twice in a year and expect to again. That is the actual lesson: nobody's stack should be load-bearing on a specific model being ahead.

How to check this yourself

This post will go stale, faster than most. Four Opus releases and two GPT generations landed in seven months, so treat any dated comparison as a snapshot.

Go to the vendor's own model or pricing page rather than an article, including this one. Both publish current lineups and rates, and both keep older announcement posts online, which is how obsolete figures stay in circulation: the announcement is a snapshot of launch day while the pricing page is current.

Check whether what you are reading names a model that still exists. The fastest credibility test on any AI comparison in 2026 is whether the models in its title are current. If it says GPT-5 or Claude 4, it was not checked this month.

And be suspicious of precise benchmark numbers without a link to the run. One decimal place is not evidence.

Verified August 4, 2026 against Anthropic's published model reference and OpenAI's GPT-5.6 announcement, with the Anthropic prices re-checked on September 1, 2026, which is when the Sonnet 5 row was corrected. Model lineups, release dates, API prices and context windows are from those sources. Consumer plan prices are sticker rates and worth confirming at checkout, since regional pricing is now common. Head-to-head capability claims are deliberately absent, as explained above.