GPT-5.6 SOL is HERE

AI summaries can be incomplete or wrong. Verify anything important against the original video.

Matthew Berman reviews GPT-5.6 Sol, a new model from OpenAI, comparing its performance and cost against Fable 5. He finds GPT-5.6 Sol to be superior in benchmarks, reasoning, and general capabilities, particularly for coding tasks, while also being more affordable.

GPT-5.6 Sol is a significant improvement over previous models, offering better performance and cost-efficiency, though its raw capabilities remain within the established boundaries.

• Fable 5

  GPT-5.6 generally performs better, especially on reasoning tasks, and is more affordable.

• GPT-5.6 Sol Performance

  GPT-5.6 Sol is a significant step forward, outperforming previous models and competitors in key benchmarks.

  • 28:10 It beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost.
  • 34:10 GPT-5.6 Sol with max reasoning comes within one point of Fable 5 while completing tasks in 61% less time at roughly half the estimated cost.
• GPT-5.6 Sol Cost Efficiency

  GPT-5.6 Sol offers a lower cost model with competitive performance, making it more affordable.

  • 28:10 It beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost.
  • 34:10 GPT-5.6 Sol with max reasoning comes within one point of Fable 5 while completing tasks in 61% less time at roughly half the estimated cost.
  • 45:50 GPT-5.6 Sol is priced per 1M tokens across three model sizes: Sol $5 input / $30 output, Terra $2.50 input / $15 output, and Luna is $1 input / $6 output.
• GPT-5.6 Sol Capabilities

  GPT-5.6 models are more capable than earlier models in both biology and cybersecurity but do not cross the Critical threshold in either category.

  • 55:00 In cybersecurity, our testing suggests GPT-5.6 is better at finding and fixing vulnerabilities than at reliably carrying out autonomous, end-to-end attacks against hardened targets—giving defenders an opportunity to strengthen systems before weaknesses are exploited.
  • 57:50 In biology, our testing suggests GPT-5.6 can support legitimate research but does not provide the end-to-end capability needed to create, engineer, or synthesize a highly dangerous novel threat.
• Codex Integration

  Codex is merging into the ChatGPT app.

  • 6:45 I switched back to GPT-5.5 to check. It went straight to the command line, even when the browser made more sense.

• 2:05 GPT-5.6 Sol is available now and is already making waves, costing less per token than previous models.
• 2:10 The government approval held it up for two weeks, and it lands head-on against Claude Fable.
• 3:00 Berman has used GPT-5.6 Sol internally for the past two months and has probably burned through more than 25 billion tokens.
• 4:00 Codex has gotten much faster lately and is very noticeable.
• 4:55 Cerebras is expected to soon serve full GPT-5.6, not a watered-down version.
• 6:55 GPT-5.6 usually takes the short route to task completion, knows how to do things really well and directly.
• 7:30 The Minecraft-style game is the best example, it started with /goal and walked away.
• 8:20 The model kept adding levels, mobs, and tiny details.
• 8:25 This model keeps adding levels, mods, and tiny details.
• 10:20 The prompt stops sooner and is more likely to ask what comes next.
• 10:40 The Codex browser is the first version of that idea I want to use all day.
• 11:10 GPT-5.6 is better at long-sustained work.
• 14:00 It keeps finding useful work when you give it a concrete finish line.
• 14:10 It keeps finding useful work when you give it a concrete finish line.
• 19:10 GPT-5.6 is faster than other models in two different ways: The raw generation speed is higher, something OpenAI has been putting effort into.
• 20:10 It also takes a shorter path to solutions. It wonders less, changes less code, and generally knows how to get things done directly.
• 22:10 GPT-5.6 Sol is capable of performing tasks faster than Fable, with roughly 3x fewer tokens per second.
• 25:00 GPT-5.6 outperforms previous and competing frontier models with fewer tokens and at lower estimated cost.
• 28:10 It beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost.
• 34:10 GPT-5.6 Sol with max reasoning comes within one point of Fable 5 while completing tasks in 61% less time at roughly half the estimated cost.
• 45:50 GPT-5.6 is priced per 1M tokens across three model sizes: Sol $5 input / $30 output, Terra $2.50 input / $15 output, and Luna is $1 input / $6 output.
• 48:30 GPT-5.6 Sol also introduces more predictable prompt caching, including support for explicit cache breakpoints and a 30-minute minimum cache life.
• 50:00 For GPT-5.6 and later models, cache writes are billed at 1.25x the model's uncached input rate, while cache reads continue to receive the 90% cached-input discount.
• 55:00 GPT-5.6 is better at finding and fixing vulnerabilities than at reliably carrying out autonomous, end-to-end attacks against hardened targets