Archived review of GPT-5's August 2025 launch: the unified-router design, the 4o backlash, what improved β and how it reads from mid-2026.
π 2026 update: This is an archived review of GPT-5, launched August 7, 2025. The line has moved on β GPT-5.4 was retired in July 2026, GPT-5.5 Instant is today's default and GPT-5.6 is rolling out. For the current state, see our ChatGPT updates tracker. We keep this review live because GPT-5's launch β technically solid, communicationally rocky β set the template for how the entire 5.x era works.
A real capability step β unified fast-response and deep-reasoning modes behind a real-time router β undercut by a bumpy rollout that OpenAI itself acknowledged.
The consensus settled where the early skeptics stood: a meaningful improvement over GPT-4, not the transformative leap the marketing implied. In head-to-heads against Claude's Opus 4.6/4.7 generation it won some categories and lost others. Its most durable legacy isn't a benchmark β it's the release-and-retire cadence and router architecture that now define OpenAI's line.
The GPT-5 model is retired. The GPT-5 architecture β router, fast/thinking split, user-selectable modes β is what every subsequent release ships on. When a launch is remembered for its structure rather than its scores, the structure was the product.
Launched August 7, 2025, GPT-5 arrived as a unified system: a fast model for most queries, a deeper reasoning model ("GPT-5 thinking") for hard problems, and a real-time router deciding which to use based on complexity, tools and intent β merging the o-series' reasoning with the GPT series' speed. It rolled out to everyone, free tier included (with delays).
Before GPT-5, choosing between speed and reasoning was the user's job, made at model-selection time. The router moved that decision to runtime and made it per-query. Every vendor now does some version of this β Google's thinking budget, Anthropic's extended thinking, OpenAI's reasoning effort.
Three self-inflicted wounds defined week one:
Sam Altman's own words: the debut was "a little more bumpy than we hoped for." Credit where due β the fixes (restored models, explicit modes) shipped fast and stuck: today's model picker exists because the unified dream didn't survive contact with users.
The GPT-5.6 family reached GA on July 9, 2026 after a limited preview from June 26, shipping as three named tiers rather than one router-mediated model:
| Model | Role | Input / Output per 1M |
|---|---|---|
| Sol | Flagship | $5 / $30 |
| Terra | Balanced, everyday work | $2.50 / $15 |
| Luna | Most cost-efficient | $1 / $6 |
On July 30, 2026, OpenAI cut Luna by 80% and Terra by 20%. Read that alongside Google pushing frontier-adjacent capability into a cheap Flash tier and Anthropic's mid-tier matching its previous flagship: the entire industry is competing at the bottom of its own price ladder.
GPT-5 tried to hide tier selection behind a router and users rejected it. GPT-5.6 ships three explicitly named tiers with published prices. The 2025 lesson β routing without transparency burns trust β is visible directly in the 2026 product structure.
Beyond the drama, the capability gains were real: stronger coding and multi-step reasoning, better tool use, and the thinking-mode foundation that later versions refined. The "bland" critique also softened over the 5.x iterations β though deliberate prompting still beats defaults for distinctive voice.
The early complaint that GPT-5 was "less creative, more corporate" wasn't wrong β it was a description of a model tuned for reliability over surprise. That trade held across the 5.x line and is now a deliberate positioning rather than a flaw, which is why prompting for voice matters more than it did in the GPT-4 era.
GPT-5 itself is retired; its architecture won. The router (now with user override), the fast/thinking split, the free-tier thinking allowance β all 5.x standards trace here. The launch also taught the industry's now-standard lesson, twice over: model attachment is real (4o), and routing without transparency burns trust (the router walk-back).
Model attachment is real. The 4o removal proved users form working relationships with specific model versions, and that deprecation is a product decision, not a maintenance one.
Routing without transparency burns trust. Users would rather choose a cheaper tier than suspect they were given one. Every vendor's 2026 pricing page reflects this.
If you're choosing a model today, no β start from the current lineup. If you're studying how AI products fail and recover in public, GPT-5's first month remains the best case study in the business.
Sources: OpenAI β Introducing GPT-5 Β· TechCrunch launch coverage Β· MIT Technology Review analysis Β· SAN on the mixed reception Β· 2026 retrospective assessment
Originally published August 2025 Β· Substantially revised July 29, 2026
Seedance 2.5 brings 30-second audio-video generation, 50 mixed references and timed editing. Here is what is official, priced and still worth testing.
MiniMax H3 generates 2K video with native stereo audio, publishes per-second pricing, and offers H3-Base weights under a Community License.
Hermes Agent blends persistent memory with self-improving skills, but its security and outcome depend on the backend, approvals, and review you choose.
OpenClaw is a high-capability agent runtime whose value depends on deliberate gateway security, skill review, and operating discipline.
AI music in 2026: Suno v5.5's capabilities and legal defiance, Udio's settlement path, the Warner deal retiring unlicensed models, and creator rules.
The AI video market after Sora's exit β with real pricing: $0.50 to $2.50 per 10-second clip, who leads the rankings now, and how to choose per shot.
The honest 2026 free-tier comparison: Gemini's generous bundle, Claude's quality-first plan, ChatGPT's capped breadth, and when free stops being enough.
An honest 2026 assessment: what GPT-5.5, Dreaming V3 memory and agents actually deliver, where Claude and Gemini beat it, and who should pay for Plus.
