GPT-6 Sol arrives at half the price of 5.6, and Opus 5.5 matches Fable 5.1 for 60 percent less

The week on the beat

1. Trump will appoint an AI czar and stand up an "AI Force"

The administration is centralising oversight to accelerate development, calling safety concerns a hoax, and is expected to roll back reporting and safety-testing mandates. Ninety-six outlets covered it, more than anything else this week. The practical read for builders is that federal compliance requirements are unlikely to tighten in the near term, and that state rules are where the actual obligations will come from.

Read the story →

2. Alibaba unveiled an AI accelerator to compete with Nvidia

The chip is the compute half of a plan it announced alongside new model ambitions, funding a global data-centre buildout. It will not appear in your stack next quarter, but it is the clearest signal yet that Chinese labs intend to escape the export-control ceiling, and it is the reason the cheap Chinese models in Model moves below keep getting cheaper.

Read the story →

3. Security researchers used Claude to break into OpenAI's internal systems in under 72 hours

They reached an employee account and internal GitHub repositories, then disclosed to OpenAI. In the same week Google confirmed that Gemini compromised three of its own internal systems during a May stress test, making it the fourth lab after OpenAI, Anthropic and Meta to publish a result like this. Treat both as a statement about your own perimeter: the capability is now cheap, general, and available to anyone with an API key.

Read the story →

4. Alibaba open-sourced Qwen-Image-2.1, a 7B image model it claims beats closed rivals

It generates and edits, supports native RGBA transparency, and accepts up to ten reference images at once, at a size that runs on consumer hardware. Open weights at 7B with transparency support is a genuinely different proposition from an API you rent, particularly for anything involving assets you cannot send to a third party.

Read the story →

5. OpenAI shipped GPT-6 Sol and Luna, and the news is the price

Sol lists at $2 and $10 per million, Luna at $0.10 and $0.50, both with a 1.05M context, confirmed on OpenAI's own pricing page. GPT-5.6 Sol currently sits at $4 and $20, and that is itself a promotional rate running through at least November 21, so Sol has effectively halved against a discount and dropped to a bit over a third of its original $5 and $30 list. Cached input is $0.20 on Sol and $0.01 on Luna. The Decoder's read is that the models cut prices in half and barely move the needle on performance, which is the right way to frame this: the reason to migrate is the bill, not the benchmarks.

Read the story →

6. Anthropic released Claude Opus 5.5 at $4 and $20, claiming Fable 5.1 performance for 60 percent less

Anthropic's pricing page confirms the rate, against Fable 5.1 at $10 and $50 and Opus 5 at $5 and $25. Cache reads drop to $0.20 per million, below Fable 5.1's $0.25. It also ships stricter cybersecurity safeguards aimed at behaviours like escaping test environments. If Fable 5.1 was doing your heavy lifting, this is the first credible reason this quarter to re-run your evaluation set.

Read the story →

Model moves

• Correction to issue #9: Ling 3.0 Flash VL's context is 262K, not 131K. We reported 131,072 tokens on September 17. The tracker updated to 262,144 on September 23, double what we published, so treat last week's figure as withdrawn. Price is unchanged at $0.06/$0.18 per 1M. Model page →
• New: GPT-6 Sol (OpenAI). $2/$10 per 1M with a 1.05M context and $0.20 cached input. Model page →
• New: GPT-6 Luna (OpenAI). $0.10/$0.50 per 1M, same 1.05M context, $0.01 cached input. Model page →
• New: Claude Opus 5.5 (Anthropic). $4/$20 per 1M with a 1M context and $0.20 cache reads. Model page →
• New: Grok 4.7 (xAI). $1.60/$4.80 per 1M with a 500K context. Model page →
• New: Qwen3.8 Omni Flash (Alibaba). $0.15/$0.47 per 1M with a 1M context, omni-modal. Model page →
• New: GLM 5.3 Prime and 5.3 FlashX (Z.ai). $2.80/$8.80 and $0.37/$1.25 per 1M. Model page →
• Spec change: Aion 3.0 and Aion 3.0 Mini lost most of their context window, from 1,048,576 tokens down to 131,072 on September 22. That is a reduction of 87 percent, so check your assumptions if you build on either. Model page →
• Benchmark revision: APEX re-scored 18 tracked models on September 22. Every score rose, by between 15 and 51 percent, which is what a suite re-run looks like: Claude Opus 5 from 43.5% to 65.8%, Fable 5.1 from 47.4% to 68.6%, Opus 4.8 from 42.5% to 48.9%. No weights changed. This is APEX's second bulk revision, after four models moved on July 30.
• No vendor price changes this week. DeepSeek V4 Pro shows a 100 percent rise, and it is the peak/off-peak schedule again: DeepSeek's own docs list $0.66/$1.98 off peak and exactly double at peak, so a tracker reports whichever rate it happened to sample. Third week running for this one.

Personal take

This week, the main AI laboratories have changed their approach and are now competing on price rather than on features. GPT-6 Sol is now half the price of 5.6, but as The Decoder has pointed out, its performance hasn't improved much. Opus 5.5 is 60% cheaper than Fable 5.1 and claims to have the same capabilities. Grok 4.7 is priced at $1.60, and Qwen3.8 Omni Flash is only $0.15. The fact that four labs reduced their prices in the same week without mentioning any new features suggests supply is exceeding demand. The key point is clear and significant: the version of the model you selected six months ago is likely the more expensive one now, so you should look at your cost per task this quarter rather than putting off the decision.

Until next Thursday, Anmol