Ever since GPT-4 launched in 2023, every month brings a fresh rumour about its successor. What we actually know as of mid-March 2026 fits in three lines: OpenAI is preparing a natively multimodal model with more stable structured reasoning and a far larger context window. Everything else deserves to be handled with care.
What OpenAI has actually said
Sam Altman has confirmed on several occasions that a model «significantly better» than GPT-4 is in training. OpenAI’s leadership describes a qualitative jump comparable to GPT-3 → GPT-4: not a simple iteration, but a step change.
Public communications remain cautious. No benchmarks, no release date, no official name — «GPT-5» is used by convention. OpenAI has, however, raised its safety testing framework, which suggests a model powerful enough to trigger the full internal protocol.
“The first time you use it, you’ll have the same kind of moment as going from GPT-3 to GPT-4.”
What the leaks suggest
Three technical clues keep coming up in infrastructure conversations:
- 01
Large-scale Mixture of Experts
Like Claude 3.5 and Gemini 1.5, GPT-5 is said to be an MoE — several specialised sub-models activated on demand, which allows the parameter count to grow without blowing up inference costs.
- 02
Native multimodality
Text, image, audio and video in and out within a single model, with no external pipeline. Reasoning carries across modalities without degradation, which changes everything for agents.
- 03
Built-in reasoning
The «thinking» mode of GPT-o1/o3 would become the default, with an adjustable cost. The model decides how much chain of thought to allocate to each question.
What GPT-5 should be able to do
If the trajectory of Claude 3.7 and Gemini 2 is any guide, we can safely anticipate:
🧠Structured reasoning
Maths, multi-step planning, complex code — without hallucinating.
🪟Stable long context
Working across 500 pages of documents with no quality drop in the middle.
🎨Natively multimodal
Image generation and video understanding in the same session.
🛠️Reliable tool use
Better tool selection, API planning, smart retries.
⚡Adjustable latency
Fast mode for chat, slow and deep mode for reasoning.
🔒Finer refusals
Fewer false positives, better grasp of business context.
How much will GPT-5 cost?
OpenAI has lowered its price per token with every major generation. The most likely trajectory:
| Model | Input / 1M | Output / 1M |
|---|---|---|
| GPT-4 (2023) | $30 | $60 |
| GPT-4o (2024) | $5 | $15 |
| GPT-4.1 (2025) | $2 | $8 |
| GPT-5 (rumoured) | ~ $5–10 | ~ $20–40 |
Batch mode and enterprise pricing should continue to offer 30 to 50 % discounts.
Newsletter
Get the AI analyses, once a month
Breakdowns like this one, no hype, no spam.
Up against Claude 4 and Gemini 3
GPT-5 is not arriving in an empty market. Anthropic and Google have raised the bar.
Anthropic
Claude 4
The benchmark for long-form reasoning and writing quality in French. Very strong on code.
Gemini 3
An edge on ultra-long context (10M+ tokens) and Workspace integration.
OpenAI
GPT-5
Has to hold its rank on versatility and win back ground on reasoning.
When should we expect it?
No confirmed date. OpenAI’s pattern suggests a rollout in several waves:
- 1
Phase 1 — Red team and partners
A few weeks of closed testing (Microsoft, governments, key enterprise partners).
- 2
Phase 2 — ChatGPT Plus and Team
Gradual rollout to paying subscribers, in regional waves.
- 3
Phase 3 — Public API
Opening to developers, with quotas and rate limits at first.
- 4
Phase 4 — Consumer pricing
Price stabilisation and arrival in the free, education and non-profit plans.
Frequently asked questions
Has GPT-5 been officially announced? +
OpenAI has confirmed it is working on a next-generation model without publishing its technical specifications. Statements from Sam Altman and Mira Murati point to a jump comparable to GPT-3 → GPT-4, with no firm timeline to date.
What do we know about the context window? +
Recent GPT-4.x rollouts already reach 128k to 1M tokens depending on the variant. Leaks point to a standard 1M-token window for GPT-5 and an extended mode for enterprises. To be confirmed at release.
Will GPT-5 be natively multimodal? +
The strong assumption is yes: text, image, audio and video in and out within a single model, with more stable reasoning from one modality to the next — in line with GPT-4o.
Will prices go down or up? +
OpenAI has a clear tendency to lower the price per token with each generation (GPT-4 → GPT-4o → GPT-4.1 all went down). Expect a GPT-5 that is more expensive in absolute terms but cheaper at equal quality, especially via batch mode.
When will GPT-5 be released? +
No confirmed date. Rumours point to a 2026 window with gradual access (enterprises first, then ChatGPT, API last).
Also worth reading
⚠️ Transparency: some links may be affiliate links. No impact on our ratings, and none on prices.