OpenAI GPT-5 Release: What It Can Do, How It Differs From GPT-4o

GPT-5 released on August 7, 2025 priced at $1.25 per million input tokens — roughly half of GPT-4o's $2.50 rate — while adding an automatic router that decides, prompt by prompt, whether to answer fast or think harder. That single architectural change, not a raw intelligence jump, is the real story of the GPT-5 release.

Almost a year later, in July 2026, GPT-5 is no longer OpenAI's frontier model — GPT-5.4, GPT-5.5, and the newly launched GPT-5.6 family (Sol, Terra, Luna) have since shipped. But GPT-5 is still the model most people mean when they compare "old ChatGPT" to "new ChatGPT," it's still the free-tier default for a huge share of ChatGPT's 900 million-plus weekly users, and it's still the baseline OpenAI itself uses to justify every model that came after it. Here's what actually changed, what it costs, and who should still be running it.

What Was the OpenAI GPT-5 Release and When Did It Happen?

OpenAI released GPT-5 on August 7, 2025, replacing the confusing model-picker menu in ChatGPT (o3, o4-mini, GPT-4.5, GPT-4o, and others) with a single unified system. Instead of asking users to guess which model fit their task, GPT-5 uses an internal router that automatically classifies each prompt and sends it down a fast path for simple requests or a slower, more deliberate reasoning path for complex ones.

The release also came with a meaningful price cut: GPT-5 launched at $1.25 per million input tokens and $10 per million output tokens, versus GPT-4o's $2.50 input rate. For API customers running high-volume workloads, that's a straightforward cost reduction on the input side, even before accounting for GPT-5's router steering easy queries toward cheaper compute automatically.

GPT-5 vs GPT-4o: Pricing and Benchmark Comparison

Figures are 2025-2026 data blended from OpenAI's developer pricing pages, Artificial Analysis Intelligence Index tracking, and per-model benchmark reporting via DocsBot and pricepertoken.com. Context windows rounded to nearest standard tier.

How GPT-5 Actually Differs From GPT-4o

The headline difference isn't raw intelligence — it's architecture. GPT-4o was a single model that answered every prompt at the same depth, whether you asked it to summarize an email or debug a distributed system. GPT-5 introduced a router that inspects each query and decides, automatically, whether to respond immediately or invoke a slower internal reasoning process. That means the same GPT-5 endpoint can feel instant for simple tasks and noticeably more deliberate for hard ones, without the user ever picking a model.

On raw capability, GPT-5 posted an MMLU Pro score of 80.6 (70th percentile among tracked frontier models at release), a GPQA score of 67.3, and a LiveCodeBench score of 55.8 — all meaningful steps up from GPT-4o's prior-generation numbers. OpenAI also reported a significant reduction in hallucination rates for GPT-5 relative to GPT-4o, particularly in long, multi-step reasoning chains where GPT-4o was more prone to confidently stating incorrect intermediate steps.

Context window matters too: GPT-4o topped out at 128K tokens, while GPT-5 expanded to roughly 400K, letting it hold far more of a codebase, contract, or document set in a single pass. That gap has since widened further — GPT-5.4 and everything after it moved to a full 1M-token window, which we cover in more depth in our context window economics post.

ChatGPT Usage and OpenAI Revenue Since the GPT-5 Release

The GPT-5 release landed in the middle of a steep usage curve. ChatGPT had roughly 400 million weekly active users in February 2025, before GPT-5 shipped; that grew to 700 million by July 2025 and crossed 900 million by March 2026. In June 2026, ChatGPT's app crossed 1 billion monthly active users — the fastest any consumer app has hit that mark, ahead of TikTok, Instagram, and Google Maps.

OpenAI's business scaled alongside that usage: the company crossed $25 billion in annualized revenue in February 2026, generating roughly $2 billion per month, with a stated full-year 2026 revenue target of $30 billion. OpenAI also passed 1 million business customers in November 2025, and by mid-2026 roughly 92% of Fortune 500 companies reported using ChatGPT in some capacity — context we track alongside other AI infrastructure spend on our AI valuations dashboard.

Who GPT-5 Is Actually For in 2026

GPT-5 is still relevant for three specific groups, even a year after its release and well after OpenAI shipped GPT-5.4, GPT-5.5, and GPT-5.6. First, free and Plus-tier ChatGPT users, where GPT-5 (or a close variant) remains the default model for a large share of the 900 million-plus weekly user base — most consumer usage never touches the newer paid-tier flagship models at all. Second, enterprise teams with evaluation suites already calibrated against GPT-5's behavior; OpenAI explicitly kept GPT-5 and the GPT-5.4 family on its live pricing sheet for exactly this reason, so teams whose evals pass on GPT-5 don't have to pay 2026-flagship GPT-5.6 rates just to stay current.

Third, cost-sensitive, high-volume API workloads that don't need frontier reasoning — classification, extraction, basic summarization — where GPT-5's $1.25 input rate is still competitive even against some of OpenAI's own newer, cheaper tiers like GPT-5.6 Luna at $1.00/M. For anyone building agentic workflows, long-context document analysis, or frontier coding tools, though, GPT-5.5 or the GPT-5.6 family is the more capable choice as of mid-2026 — GPT-5's 400K context window and 55.8 LiveCodeBench score are now roughly a generation behind GPT-5.4's 1M-token window and 75% computer-use benchmark.

If you're deciding which model to build on today, treat GPT-5 the way most infrastructure teams treat any N-1 release: stable, well-understood, and cheaper than the frontier tier, but not the model to pick if the task genuinely needs OpenAI's best available reasoning. We track how these pricing shifts ripple through AI infrastructure spend more broadly on our big tech earnings tracker.

There's also a migration-cost argument for staying on GPT-5 longer than the benchmark charts suggest you should. Every model swap means re-running eval suites, re-tuning prompts, and re-checking latency and cost assumptions across a production pipeline — work that isn't free even when the newer model is objectively better. Teams running dozens of GPT-5-tuned prompts in production, especially in regulated industries like fintech or healthcare where prompt behavior gets audited, often rationally choose to absorb a capability gap rather than re-certify an entire pipeline against GPT-5.6 every time OpenAI ships a new tier.

What Came After the GPT-5 Release: The Full 2026 Timeline

OpenAI's post-GPT-5 release cadence has been fast even by its own standards. GPT-5.4 shipped March 5, 2026 with a 1M-token context window, 75% on computer-use benchmarks, and pricing of $2.50/$15 per million tokens — effectively GPT-4o-level pricing for a much more capable model. GPT-5.5 followed on April 23, 2026 at $5/$30 per million tokens, posting an 82.7% Terminal-Bench score and establishing itself as the frontier coding model for several months.

Most recently, OpenAI launched the GPT-5.6 family on July 9, 2026 — Sol ($5/$30, frontier capability), Terra ($2.50/$15, balanced production work), and Luna ($1/$6, cost-sensitive volume) — replacing the single-flagship model with a three-tier lineup that mirrors how enterprises actually route traffic across cost and capability needs. That progression, from one model at release to a segmented three-tier family less than a year later, is itself the clearest signal of how fast the GPT-5-era pricing and capability curve has moved.