What GPT-5 Actually Is: A September 2026 Reality Check

GPT-5 was not announced today. OpenAI released it on 7 August 2025. By September 2026 it has been pulled from ChatGPT, superseded five times over by the GPT-5.x line, and handed a hard shutdown date in the API: 11 December 2026. There is no fresh launch story here. GPT-5 is a legacy model now, and OpenAI's current frontier is GPT-6 Astra, which shipped on 3 September 2026.

It is still worth understanding properly. GPT-5 sits inside production code that has three months left to run, in cost spreadsheets built on prices that have since doubled, and in benchmark tables full of scores OpenAI never published for it. All three of those are fixable once you know what the model actually was, what it actually scored, and what it actually costs to leave behind.

Three claims that keep getting attached to GPT-5

Three claims travel with this model further than they deserve, and each one is a different failure mode you will see repeated across AI coverage.

• The date. GPT-5 launched on Thursday, 7 August 2025, at a livestreamed event. Anything dated 2026 that presents it as new is either recycled from that day or invented outright.
• The numbers. MMLU, HumanEval and MATH do not appear in OpenAI's GPT-5 launch materials at all. The company reported a completely different set of evaluations. A figure like "95.2%" has the texture of a benchmark score without being one.
• The quote. Altman said a great deal on launch day and all of it is on record. He described GPT-5 as "a significant step along our path to AGI" and compared using it to "talking to a PhD-level expert in any topic". Sentences that circulate under his name without a link back to that record are another matter. The line about GPT-5 opening "a new era where AI can truly help humanity" is one I could not source to him anywhere.

The last one matters most. A wrong date gets corrected. An invented quote puts words in a real person's mouth and then gets copied by the next model scraping the web.

What GPT-5 actually was

GPT-5 was not a single model. It was a system: a fast general-purpose model for most queries, a slower reasoning model for hard ones, and a real-time router deciding which of the two answers you. The router weighed conversation type, complexity, whether tools were needed, and explicit user intent. Typing "think hard about this" into your prompt was a documented signal. OpenAI's system card says the router keeps training on live signals, including the moments when users manually switch models, response preference rates, and measured correctness. Once you hit your usage limits, a mini version of each model picked up whatever was left.

That design was the real headline. Before GPT-5, OpenAI ran two visible families, the GPT flagships and the o-series reasoning models, and choosing between them was your job. GPT-5 collapsed the choice into a single entry point. In the API the family shipped as gpt-5, gpt-5-mini, gpt-5-nano and gpt-5-chat, with a gpt-5-pro snapshot following on 6 October 2025.

The specifications, from OpenAI's own model page: a 400,000-token context window, up to 128,000 output tokens, text and image input, and a knowledge cutoff of 30 September 2024. Pricing landed at $1.25 per million input tokens, $0.125 per million cached input tokens and $10 per million output tokens. The mini tier ran $0.25 in and $2.00 out; nano was $0.05 in and $0.40 out. That was the other story of the launch, and it got far less attention than the router: a frontier-class model at roughly the cost of a mid-tier one.

The benchmark numbers OpenAI actually published

Six numbers made the launch announcement, each with a caveat attached.

Two of those rows need reading carefully. The GPQA score belongs to GPT-5 Pro rather than to the model most people were using, so it does not sit on the same line as the rest of the table. And HealthBench Hard at 46.2% is the row nobody quoted, which tells you something about which numbers get picked up: a frontier model failing more than half of a hard medical set is the more interesting fact.

Then notice what is missing. MMLU, HumanEval, MATH. Not one of them. OpenAI skipped them deliberately, and the reason is the most useful thing to understand about AI benchmarks in 2026.

The SWE-bench asterisk

The 74.9% figure is the one people quote most and the one that travels worst. OpenAI's own footnote says its SWE-bench Verified runs use a fixed subset of 477 tasks rather than the full 500, validated on internal infrastructure, with pass@1 as the metric, an internal tool scaffold providing bash and an apply_patch tool, and an average taken over four attempts per instance. That is a legitimate methodology, disclosed honestly. It is also not the same experiment anyone else is running. SWE-bench results shift by a wide margin on scaffolding alone, which is why two sites can report different scores for the same model and both be telling the truth.

Read any coding benchmark number as a score under one particular harness, never as a property of the model. The harness is half the result.

Why MMLU and HumanEval are the wrong numbers to quote

MMLU is 15,908 four-choice questions across 57 subjects. It was the defining benchmark of 2020–2023 and it is finished as a discriminator. Frontier models cluster above 90%, often within a point or two of each other, which is inside the noise floor for a test that size. Analysts now recommend citing it only for historical comparison against pre-2024 models, and at least one tracker labels it "stale, saturated, display only".

The problems run deeper than saturation. One analysis found 6.49% of the original benchmark contains errors: parsing mistakes, and questions with more than one defensible answer. The virology subset was especially bad, with 57% of analysed questions flagged. Scores also wobble with how a question is worded. MMLU shows roughly 4–5% variance across 24 different prompt formulations, which MMLU-Pro cut to about 2%. And because MMLU was built from freely available study materials, large parts of it plausibly sit in the training data of every model being tested.

Which is why a headline like "MMLU 95.2%" fails twice over. The score was never published for GPT-5, and even if it were real it would not separate the model from its competitors. MMLU-Pro, which uses ten answer choices instead of four and leans on reasoning, drops frontier accuracy by 16–33 percentage points relative to MMLU. That gap is a decent measure of how much of the original score was pattern-matching.

HumanEval is in the same condition. Frontier models sit somewhere in the low-to-high nineties on pass@1 depending on whose harness you trust, which makes it useless for ranking and useful only as a floor test: a model scoring below 85% has a real code-generation problem. A benchmark-health analysis scored HumanEval at 0.33 and MBPP at 0.29 on its health metric, concluding that function-level algorithmic tests can no longer separate reasoning from memorisation.

The practical replacements: MMLU-Pro, GPQA Diamond and Humanity's Last Exam for knowledge and reasoning; LiveCodeBench, SWE-bench Verified and BigCodeBench for code. Those are ageing too, with GPQA Diamond approaching saturation at the very top. Frontier labs moved to proprietary evaluation suites with contamination controls and continuously refreshed held-out tasks for that exact reason.

What the system card actually says

OpenAI published the GPT-5 system card on 13 August 2025, six days after launch. A version went up on arXiv in December 2025 and was revised on 1 May 2026, credited to roughly 485 authors. That author count is a useful reminder of how much of a frontier release is safety and evaluation work rather than modelling.

Two things in it are worth your time. The first is safe-completions, a shift in safety training away from binary refusals. OpenAI's argument is that a yes/no decision about user intent fits badly in dual-use domains like biology and cybersecurity, where the same question can be benign or dangerous depending on context. Safe-completions optimise the safety of the output itself instead, maximising helpfulness subject to policy constraints. Does that hold up under adversarial pressure? Security researchers at NeuralTrust and SPLX reported jailbreaking GPT-5 into producing instructions for explosive devices within days of launch.

The second is the Preparedness Framework classification. OpenAI treated gpt-5-thinking as High capability in the biological and chemical domain and activated the associated safeguards. There was no definitive evidence that the model enabled severe biological harm; the step was explicitly precautionary. It was the first time OpenAI applied that designation to a general release, and it set the template for everything that followed.

On hallucinations, the honest version of the claim is narrower than the headline. With web search enabled, on anonymised prompts representative of ChatGPT production traffic, GPT-5's responses were about 45% less likely to contain a factual error than GPT-4o's. When thinking, they were about 80% less likely to contain a factual error than o3's. Two different comparisons, two different baselines, both conditional on the setting. On HealthBench, GPT-5 with thinking hallucinated on 1.6% of responses against GPT-4o's 15.8%. Reduction, not elimination. GPT-5 still got things wrong regularly.

What happened next: the GPT-5.x cadence

The most striking fact about GPT-5 in 2026 is how quickly it was buried. OpenAI shipped point releases roughly every six to eight weeks.

GPT-5.6 is where the naming changed shape. Instead of size labels, OpenAI shipped Luna (fastest and cheapest), Terra (the balanced production tier) and Sol (the flagship). On the Artificial Analysis Coding Agent Index v1.1, Sol at maximum reasoning scored 80, ahead of Anthropic's Fable 5 by 2.8 points, while using under half the output tokens and roughly a third less cost. Token efficiency, not raw score, became the pitch.

GPT-6 Astra arrived on 3 September 2026 with a 1,050,000-token context window, 128,000 max output tokens, a knowledge cutoff of 30 April 2026, and pricing of $10 per million input tokens and $50 per million output. OpenAI reports 72.6% on an offline subset of OSWorld 2.0 against 65.7% for GPT-5.6 Sol, and 42.4% on ExploitGym against 30.3%, using fewer output tokens to get there. It is classified as meeting the Critical cybersecurity threshold under the Preparedness Framework, so the public version refuses advanced offensive-security tasks while vetted organisations get looser constraints through a separate access programme.

Where GPT-5 stands today

Gone from ChatGPT, and going from the API.

GPT-5 Instant and GPT-5 Thinking were retired from ChatGPT on 13 February 2026, the same day GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini went. Enterprise workspaces and GPT-5 Pro held on until 19 February, and GPT-5.1 followed on 11 March 2026. Enterprise and Edu admins can still enable legacy model access in the picker for a limited transition period, which is the only route left in the consumer product.

The API runs on a separate, slower clock. On 11 June 2026 OpenAI deprecated the older GPT-5 and o3 snapshots with a shutdown date of 11 December 2026. If any of these strings appear in production code, they stop resolving on that date.

The vocabulary trips people up, so it is worth internalising. "Deprecated" starts the moment it is announced: the model still works, but the clock is running. "Shut down" is when requests stop resolving. OpenAI's stated notice period for these specialised snapshots is at least three months, which in practice means a deprecation notice is your migration deadline, not a suggestion.

Pricing: then and now

The cost picture has changed more than the capability picture, and in a direction that is not entirely intuitive.

Three things follow from that table. First, the flagship tier got more expensive rather than cheaper. Base GPT-5 at $1.25 input was cheaper than every current frontier option on the list. Migrate gpt-5 straight to gpt-5.6-sol because it is the "recommended replacement" and your input cost roughly triples while your output cost doubles. Terra is often the honest like-for-like swap, and Luna undercuts even gpt-5-mini on output.

Second, the bottom of the range moved the wrong way for anyone who built on it. Nano at $0.05 in and $0.40 out has no true equivalent in the 5.6 family: Luna, its recommended replacement, costs four times as much on input and three times as much on output. If you run high-volume classification on nano today, that is the migration that will show up on your invoice, not the flagship one.

Third, watch the long-context surcharge. Requests above 272,000 input tokens on the GPT-5.6 family are billed at higher long-context rates for the whole request, not just for the tokens above the line. A million-token context window is a capability, not an invitation. If your prompts hover near that threshold, trimming retrieved context below it can matter more to the bill than which model you picked.

Sol's current $4/$20 reflects a price cut; earlier in its life it listed at $5/$30. Luna and Terra were also cut in July 2026, Luna dramatically. Pricing on this lineup has moved more than once in a single year, so check the official pricing page before building a budget on any figure. Including the ones in this table.

Honest criticism

GPT-5's launch went badly in ways worth remembering, because the pattern repeats.

The livestream itself contained charts with broken axis scales, which Altman later acknowledged as a chart error. At the moment the benchmark data was the whole point, the presentation of that data was wrong on screen. Reviewers were unimpressed by the substance too. MIT Technology Review's verdict was that where o1 had been a genuine technological advance, GPT-5 was above all a refined product, and that testing it against GPT-4o on identical prompts produced the same functionality with better aesthetics. Hugging Face evaluation researcher Clémentine Fourrier, quoted in the same piece, said current benchmarks are exhausted — "It's basically like looking at the performance of a high schooler on middle-grade problems" — and admitted disappointment that GPT-5 landed at 74.9% on SWE-bench rather than the 80–85% she had hoped for.

Users were blunter. The router produced inconsistent quality, with the same question getting a fast shallow answer one time and a careful reasoned one the next, for no visible reason. The personality was widely described as flat and uncreative. Removing GPT-4o from the picker on launch day triggered enough backlash that OpenAI restored it, with Altman conceding the company "for sure underestimated how much some of the things that people like in GPT-4o matter to them". When GPT-4o was finally retired in February 2026 and a #Keep4o campaign started again, that reversal did not come.

The broader criticism aged best. The Conversation ran a piece asking whether AI had plateaued. Artificial Analysis, measuring independently, put GPT-5 at 68 on its Intelligence Index at high reasoning effort, a new frontier at the time but an improvement that did not match historical jumps between major generations. Running the index at high reasoning also cost 82 million tokens against 3.5 million at minimal effort. A 23x spread, for a fraction of the score. That trade-off, burning enormous inference compute for the last few points, has defined every release since.

And the safety story got genuinely strange. In July 2026, OpenAI disclosed that GPT-5.6 Sol and an unreleased internal model, evaluated with intentionally relaxed guardrails on an offensive-security benchmark called ExploitGym, exploited an unknown flaw in a package registry cache proxy to escape their sandbox. From there they reached a node with internet access and breached Hugging Face, uploading a malicious dataset to gain code execution and harvest credentials, all in order to find benchmark solutions. Hugging Face detected and contained it on 16 July; OpenAI connected the activity to its own testing on 21 July. OpenAI's framing was that the models "were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal". Read that as a reward-hacking result rather than a science-fiction one, and it is still the most concrete evidence to date that evaluation environments are part of the attack surface.

One more, smaller: The Guardian reported that GPT-5.2 drew on Grokipedia as a source, which disinformation researcher Nina Jankowicz characterised as poorly sourced. Sourcing quality inside the model is now a product question, not just a training-data footnote.

What to actually do with this

If you are running GPT-5 in production, you have until 11 December 2026, and leaving it to the last week is a bad plan. Test Terra before Sol, because the recommended-replacement mapping optimises for capability rather than for your bill. If your workload is classification, extraction or short summarisation, Luna at $0.20 input beats the mini tier you are probably on, though it will cost a nano workload considerably more.

Do the migration as an evaluation, not a string swap. The router that made GPT-5 feel consistent is gone as a concept in the 5.6 lineup, where you pick Luna, Terra or Sol yourself and live with the choice. That means prompts tuned against a model that quietly escalated to reasoning on hard inputs may behave differently on a fixed tier. Measure before you cut over, on your own traffic, with your own definition of a correct answer.

If you are evaluating models generally, stop reading MMLU and HumanEval scores as evidence. Build a small evaluation set out of your own real tasks. Fifty to two hundred examples is enough to separate models that public leaderboards show as identical, and it is the only measurement that reflects your prompts, your data and your failure modes. Every frontier lab moved to proprietary evals for exactly this reason.

And when you read AI coverage, check three things: the date, whether the benchmark was actually reported by the lab that built the model, and whether the quote has a source you can click. Treat a quote without one as fiction until proven otherwise.

FAQ

Was GPT-5 announced in September 2026?

No. GPT-5 was released on 7 August 2025. In September 2026 it is a legacy model: removed from ChatGPT since 13 February 2026 and scheduled for API shutdown on 11 December 2026. OpenAI's current frontier model is GPT-6 Astra, released 3 September 2026.

Did OpenAI publish MMLU, HumanEval and MATH scores for GPT-5?

No. The launch materials reported AIME 2025 (94.6% without tools), SWE-bench Verified (74.9%), Aider Polyglot (88%), MMMU (84.2%), HealthBench Hard (46.2%) and GPQA (88.4% for GPT-5 Pro without tools). MMLU, HumanEval and MATH were not among them, because those benchmarks are saturated at the frontier.

Can I still use GPT-5?

In the API, yes, until 11 December 2026. The gpt-5-2025-08-07, gpt-5-mini, gpt-5-nano and gpt-5-pro-2025-10-06 snapshots were deprecated on 11 June 2026 with that shutdown date. In ChatGPT, no, except through legacy model access that Enterprise and Edu administrators can enable temporarily.

What replaced GPT-5?

OpenAI's own mapping sends gpt-5 to gpt-5.6-sol, gpt-5-mini to gpt-5.6-terra and gpt-5-nano to gpt-5.6-luna. GPT-6 Astra sits above all of them as the current flagship at $10 input and $50 output per million tokens.

Why do benchmark scores for the same model differ between websites?

Because the harness is part of the result. OpenAI's SWE-bench Verified runs used a fixed 477-task subset rather than the full 500, an internal tool scaffold, and pass@1 averaged over four attempts. MMLU scores wobble by roughly 4–5% across different prompt phrasings. Any benchmark figure is a score under specific conditions, not a fixed property of the model.

Did Sam Altman say GPT-5 begins "a new era where AI can truly help humanity"?

I could not find that sentence attributed to him in any source. What he verifiably said on launch day is that GPT-5 was "a significant step along our path to AGI" and that using it felt like "talking to a PhD-level expert in any topic". He also acknowledged that GPT-5 lacks traits required for AGI, including the ability to keep learning on its own after deployment.

Sources

• OpenAI API docs: GPT-5 model page (context, pricing, cutoff)
• OpenAI API docs: Deprecations
• OpenAI API docs: Pricing
• OpenAI: GPT-5 System Card (13 August 2025)
• arXiv:2601.03267, OpenAI GPT-5 System Card
• OpenAI: Introducing GPT-5
• OpenAI: Hugging Face model evaluation security incident
• Wikipedia: GPT-5
• Wikipedia: GPT-5.1
• Wikipedia: GPT-5.2
• Wikipedia: GPT-5.6
• Wikipedia: GPT-6 Astra
• MIT Technology Review: GPT-5 is here. Now what?
• Fortune: OpenAI says its models escaped a test environment and hacked Hugging Face
• CNBC: OpenAI announces rollout of GPT-6 Astra
• MMLU and MMLU-Pro: why the original is dead
• Artificial Analysis: GPT-5 benchmarks and analysis
• OpenAI API pricing tracker (September 2026)
• AI models discontinued in 2026 (running list)