New tools5 min readMarch 15, 2026

GPT-5: what we know (and hope for) about OpenAI's next model

Architecture, context window, reasoning, pricing: a synthesis of credible leaks, announcements and signals as of today.

Comparateur-IA

Published March 15, 2026

Every 18 months, the AI conversation pivots. GPT-5 is shaping up to be that pivot for 2026 — bigger context, native multimodality, sharper reasoning, and a new pricing curve. Here's what's credible, what's noise, and what to plan for.

1M+
context tokens

Default expected

2026
release window

Mid-year, phased

32K
output tokens

Estimated cap

01Signals

Leaks & signals to date

OpenAI hasn't confirmed timing or specs publicly, but multiple credible signals — partner briefings, hiring posts, infra leaks — point to a mid-2026 launch starting on API + ChatGPT Pro.

The headline expectation: better reasoning with significantly fewer hallucinations on long, multi-step tasks. The benchmark race has shifted from raw performance to reliability under length and complexity.

02Architecture

What's likely under the hood

Most signals point to a multimodal-native architecture with text, image and audio in one tokenizer — not bolt-on adapters. Expect tighter coupling between modalities and reduced latency on multimodal prompts.

On reasoning, the trend is "thinking models" with extended internal compute (similar to o1/o3 lineage). GPT-5 likely exposes a knob: fast/standard/deep, with prices tracking compute time.

03Context

Context window & memory

Default window: 1M tokens, with a 2M+ extended tier. That changes what's possible: full-codebase analysis, multi-document research, long-running agents without RAG hacks for many cases.

Equally important: retention quality across the window. Anthropic Claude is the reference here today. GPT-5 needs to match it, not just match the count.

Newsletter

Get AI analysis, once a month

Insights like this one — no hype, no spam.

04Pricing

Pricing & rate limits

The big unknown. Frontier models are expensive to run. Expect GPT-5 to launch at a premium over GPT-4 family, with steady price drops over 6-12 months as capacity scales.

For builders: design with model-pricing volatility in mind. Use abstractions that let you swap models per task class, and benchmark on your actual workflows — not on the OpenAI marketing graphs.

We're not waiting for GPT-5 to fix our reliability problems. We're shipping with what works today and we'll upgrade the day pricing makes sense.
Engineering lead, AI startup, March 2026
05FAQ

Frequently asked questions

When does GPT-5 launch?

OpenAI hasn't confirmed a date. Credible signals point to mid-2026 with phased rollout, starting with API and ChatGPT Pro tiers.

Will GPT-5 be multimodal natively?

Yes — text, image, audio, and likely video as input. Output is text and image at first; native video output expected later.

How big is the expected context window?

Industry signals suggest a default 1M-token window with a 2M+ extended mode. Expect higher output token limits too (~32K).

Will it kill GPT-4 and earlier?

No, but they'll be cheaper. OpenAI typically keeps prior generations as cost-optimized options for use cases that don't need frontier capability.

What's the biggest unknown?

Pricing and rate limits. Frontier models are expensive — actual unit economics will determine whether GPT-5 stays a frontier API or becomes a commodity quickly.

Related reads

⚠️ Transparency: some links may be affiliate links. No impact on our evaluations or prices.