Price and Performance on Four frontier models: Fable, Mythos, Sol, and GLM-5.1

Ronni Holmvig Strøm · 2026-06-29

Three labs shipped frontier models inside a single quarter. Zhipu AI released GLM-5.1 on March 27. Anthropic followed with Claude Fable 5 and Claude Mythos 5 on June 9. OpenAI previewed the GPT-5.6 series, led by Sol, on June 26. The models target overlapping work, mostly long-horizon reasoning, coding, and agentic tasks, and they diverge most on how a developer actually gets to them and what they cost. For anyone building durable systems on top of these, the question is no longer whether a frontier model exists for a task but which tradeoff of price, access, and capability to take.

What Each Lab Shipped

Claude Fable 5 and Claude Mythos 5 share one set of capabilities and one price. Anthropic calls Fable 5 its most capable widely released model. Mythos 5 is the same model without the safety classifiers that let Fable 5 decline a request, and it ships only in limited release through Project Glasswing. Both carry a 1M-token context window, up to 128k output tokens per request, and pricing of $10 per million input tokens and $50 per million output. Adaptive thinking is always on, and the raw chain of thought is never returned.

OpenAI's GPT-5.6 launched as a three-tier family rather than a single model. Sol is the flagship, priced at $30 per million input tokens and $60 per million output. Terra is the balanced tier at $2.50 and $15, which OpenAI positions near GPT-5.5 capability at roughly half the cost. Luna is the cheapest and fastest at $1 and $6. OpenAI names coding, science, and cybersecurity as Sol's strongest areas, and states that GPT-5.6 did not reach the critical threshold in its preparedness framework.

GLM-5.1 is the oldest of the four and the most open. It is a 744B-parameter mixture-of-experts model with 40 to 44 billion active parameters per token, a 200K context window, and MIT-licensed open weights. Zhipu trained it on 100,000 Huawei Ascend 910B chips with no Nvidia hardware. GLM-5.1 is a post-training refinement of GLM-5 focused on coding, and it is priced at $1.00 per million input tokens and $3.20 per million output. It is text-only, with no image, audio, or video input.

Price and Access Diverge Sharply

Cost separates the three flagships more than any benchmark does. On flagship input and output, GLM-5.1 runs at $1.00 and $3.20, Fable 5 at $10 and $50, and Sol at $30 and $60. Sol's output token is roughly nineteen times the price of GLM-5.1's.

Access runs in the opposite direction from price. GLM-5.1 is open-weights under an MIT license, deployable by anyone with the hardware. Fable 5 is generally available across the Claude API, AWS, Bedrock, Google Cloud, and Microsoft Foundry. Mythos 5 is restricted to approved Project Glasswing customers. GPT-5.6 is the most gated: OpenAI began with a limited preview of roughly 20 organizations at the request of the U.S. government, which cited national security concerns, and has said broader availability will follow in the coming weeks.

The Benchmark Picture, With Caveats

On coding, the published numbers cluster on a single benchmark and come mostly from the labs themselves. Fable 5 leads SWE-bench Verified at 95% and posts 80.3% on SWE-Bench Pro, ahead of Claude Opus 4.8's 69.2%. GLM-5.1 reports 77.8% on SWE-bench Verified and 52.0% on LiveCodeBench, and Zhipu claims its coding score reaches 94.6% of Claude Opus 4.6's. OpenAI has not published numeric coding benchmarks for Sol, describing its gains qualitatively.

Two caveats apply to all of it. The headline figures are largely vendor-reported, run on each lab's own scaffolding, so cross-model comparison is approximate. Independent evaluation of Fable 5's SWE-Bench Pro score and Zhipu's 94.6%-of-Opus claim was still pending at the time each was published.

The four also split on modality and architecture. GLM-5.1 is text-only and open-weight. Fable 5 supports vision and ships closed. The GPT-5.6 tiers are closed and multimodal. For anyone building on top of them, the quarter handed over three usable frontier options at three price points and three levels of openness, which is far more room to design real post-deployment systems than the field had in March.