GPT-5.6 Price Leak: iris-alpha vs Claude Fable 5
·6 min read
TL;DR — GPT-5.6 Pricing Outlook
Internal logs from OpenAI's Codex backend suggest GPT-5.6 (codename: iris-alpha) is targeting a price point roughly one-third of Claude Fable 5 — potentially around $3–4 per million input tokens if Fable 5's $10/1M is the reference. That would make it notably cheaper than the current GPT-5.5 ($5/1M input) while delivering a 43% larger context window. Nothing here is official; OpenAI has not announced GPT-5.6. All specs below come from developer log leaks and independent testing.
What Is GPT-5.6 / iris-alpha?
OpenAI developers discovered the codename iris-alpha inside Codex backend logs in June 2026 — the same channel where gpt-5.5 appeared weeks before its launch. Two related experimental models, ember-alpha and beacon-alpha, were spotted alongside it, suggesting an active evaluation cluster rather than a single isolated model.
The product roadmap inferred from the logs points to the same three-tier structure as GPT-5.5: Mini / Standard / Pro. The leaked specs below apply primarily to the Standard and Pro tiers.
⚠️ All data in this article is from leaks and third-party testing. OpenAI has not officially confirmed any of these figures. Pricing in particular is speculative.
Leaked Specifications: GPT-5.6 vs GPT-5.5 vs Claude Fable 5
| Specification | GPT-5.6 (leaked) | GPT-5.5 | Claude Fable 5 |
|---|---|---|---|
| Context window | 1,500,000 tokens | 1,050,000 tokens | 1,000,000 tokens |
| Context increase | +43% vs GPT-5.5 | — | — |
| Token efficiency | +10–15% vs GPT-5.5 | — | — |
| Knowledge cutoff | December 2025 | August 2025 | March 2025 |
| Reasoning level | xhigh (new tier) | high | extended thinking |
| Built-in CoT | Not supported | Not supported | Supported |
| Agentic coding | Reportedly beats Anthropic Mythos | Strong | Industry-leading |
| Vision / Image-to-Code | Design → runnable code | Standard | Standard |
| Browser automation | Playwright (real clicks) | Limited | Computer use |
| Internal codename | iris-alpha | — | — |
| Est. input price / 1M | ~$3–4 (estimated) | Live | Live |
GPT-5.6 pricing is estimated based on leaked positioning data ("~1/3 of Claude Fable 5"). Confirmed prices for GPT-5.5 and Claude Fable 5 are tracked live on ComputeUnion.
Pricing Analysis: Is GPT-5.6 Cheaper Than GPT-5.5?
The most striking claim in the leak is the pricing target: approximately one-third of Claude Fable 5. Working backwards from Claude Fable 5's current official rate of $10 per million input tokens and $50 per million output tokens:
| Model | Est. Input / 1M | Est. Output / 1M | vs Claude Fable 5 |
|---|---|---|---|
| Claude Fable 5 | $10.00 | $50.00 | Reference |
| GPT-5.6 (estimated) | ~$3.30 | ~$16.70 | ~1/3 of Fable 5 |
| GPT-5.5 (current) | $5.00 | $30.00 | ~1/2 of Fable 5 |
If accurate, GPT-5.6 would be cheaper than GPT-5.5 on both input and output — following the broader trend of frontier model prices falling with each generation. When GPT-5.6 launches, live pricing across OpenAI and relay platforms will be tracked at ComputeUnion.
The 1.5M Token Context Window: What It Means in Practice
The jump from 1.05M (GPT-5.5) to 1.5M tokens is a 43% increase in a single generation. At 1.5M tokens, GPT-5.6 can hold approximately:
- A 1,200-page technical book in a single prompt
- A large codebase (~40,000–60,000 lines of code)
- An entire year's worth of emails or Slack history
- Multi-day agent conversations without context truncation
Combined with the reported +10–15% token efficiency improvement, GPT-5.6 processes more content per dollar than GPT-5.5 even before the price reduction kicks in.
New Capabilities: From Playwright to 3D SVG
Agentic Coding
The most technically significant claim: GPT-5.6 reportedly outperforms Anthropic's Mythos family on agentic coding benchmarks. Pro tier tasks are said to support 20–40 minute thinking windows — longer than any publicly available model today.
Browser Automation
GPT-5.6 Pro is reported to run actual Playwright browser sessions: real clicks, form inputs, and page navigation — architecturally closer to Anthropic's Computer Use API but integrated natively into the model's agent loop.
Visual Generation
- Image to Design: Upload a screenshot or mockup → get runnable front-end code
- SVG 3D objects: Generate 3D objects in SVG format, rotatable in-browser
How to Choose: GPT-5.6 vs GPT-5.5 vs Claude Fable 5
Wait for GPT-5.6 if:
- You need the largest possible context window (1.5M vs 1M)
- Price matters and you're currently on Fable 5 or Claude Opus
- Your workflow involves agentic coding or browser automation at scale
- You're building agents that need extended thinking time (Pro tier)
Stay on GPT-5.5 if:
- You need a stable model today with known pricing
- 1M context window is sufficient
- You're optimizing for the lowest input price on read-heavy workloads
Choose Claude Fable 5 if:
- You need built-in Chain-of-Thought in the model architecture
- Enterprise compliance requirements are non-negotiable
- 1M context fits and you're not price-sensitive
Author: yego | Source: Developer log leaks, third-party testing, and independent research. All GPT-5.6 specs are unconfirmed pending official OpenAI announcement. Last updated: June 22, 2026.
Frequently Asked Questions
Is GPT-5.6 officially announced?
No. As of June 2026, OpenAI has not made any official announcement about GPT-5.6. All current information comes from developer log leaks (the iris-alpha codename) and third-party testing.
Is GPT-5.6 cheaper than GPT-5.5?
Based on leaked pricing positioning — approximately one-third of Claude Fable 5 — yes, GPT-5.6 is expected to be cheaper than GPT-5.5's current $5/1M input price. The estimated range is $3–4/1M, though the official rate won't be confirmed until launch.
Is GPT-5.6 cheaper than Claude Fable 5?
Significantly, if the leak is accurate. Claude Fable 5 is priced at $10/1M input and $50/1M output. GPT-5.6's estimated target of ~1/3 of that puts it around $3.30/$16.70 — roughly 67% cheaper on input.
What does the iris-alpha codename mean?
It's OpenAI's internal development codename, discovered in Codex backend logs. The appearance of ember-alpha and beacon-alpha alongside it suggests an active A/B evaluation cluster, not just a single prototype.
Does GPT-5.6 support thinking mode / Chain-of-Thought?
The leaked specs indicate GPT-5.6 does not have built-in CoT in the way Claude Fable 5 does. However, it supports a new 'xhigh' reasoning level setting, and Pro tier users can access 20–40 minute thinking windows for complex tasks.
When will GPT-5.6 launch?
Unknown. Based on the pattern from GPT-5.5's Codex log appearance to its public launch, the window is typically 2–6 weeks. ComputeUnion will update live pricing as soon as the API is available.