Three OpenAI models, one goal: return to the top

OpenAI launched GPT-5.6 just eleven days after the US government lifted a temporary ban on its release. The model comes in three variants: Sol (highest performance), Terra (mid-range), and Luna (affordable variant). And the very first benchmarks make it clear that OpenAI is serious about returning to the top. In the Terminal-Bench 2.1 test, which measures a model's ability to work directly in a developer terminal, Sol achieved 88.8% accuracy, narrowly surpassing the previous leader — Anthropic Claude Fable5 with 88%. In the composite Artificial Analysis Index, which combines nine diverse benchmarks, Sol scored 59 points, just one point less than Fable5. But the most interesting result comes from the so-called Agent Final Exam — a test suite of roughly 1,500 tasks that simulate real-world assignments for AI agents. Here Sol reached 52.7%, a dramatic leap compared to Fable5 (40.5%). And as if that weren't enough, the compute cost to complete the entire test suite was $1,087 for Sol — less than half the $2,315 that Fable5 consumed.

Anthropic: king of overall score, but expensive

Anthropic, with its Claude Fable5 model, currently holds the crown in the overall Artificial Analysis Index (60 points). It's a model that excels across disciplines — from programming to logical reasoning to language comprehension. But it has one fundamental weakness: price. API fees for Fable5 amount to $10 per million input tokens and $50 per million output tokens. For comparison, that's roughly ten times what Meta charges for its new model. For companies that need to process large volumes of text, Fable5 quickly becomes expensive, even though its quality remains unmatched.

Meta: when performance falls short, here comes a price bomb

Meta's development of its own AI models hasn't been going according to plan so far. Its first Muse Spark model, released three months ago, lagged so far behind in performance that Meta's market share in generative AI reached only 2.5% in May. Internally, the company faces personnel turmoil — some employees have described working conditions in the AI division as a "labor camp." Now Muse Spark 1.1 arrives, and Meta has bet everything on one card: price. In Terminal-Bench 2.1, the model scored 69.2% — not even close to Sol and Fable5. But the API costs just $1.25 to $4.25 per million tokens, roughly ten times less than Fable5. For startups, developers of smaller applications, or companies that don't need absolute top-tier performance, this is a game-changing offer. For the price of one query to Fable5, you can run ten times as many.

SpaceXAI: Grok 4.5 bets on efficiency

Elon Musk's SpaceXAI (formerly xAI) launched Grok 4.5 a day before the competition. The model placed sixth in the Artificial Analysis Index with 54 points — behind five models from OpenAI and Anthropic, but ahead of Google's Gemini 3.5 Flash. Here too, the main draw is price: API fees range between $2 and $6 per million tokens. SpaceXAI has clearly understood that it can't yet win in a head-to-head battle with Anthropic and OpenAI, so it's focusing on cost efficiency. The company also leases its compute capacity to Anthropic — an ironic detail that shows how interconnected the AI infrastructure market really is.

Google: radio silence

The biggest loser of the current wave is Google. The company last released a new model in May — Gemini 3.5 Flash — and has been silent ever since. The model has meanwhile dropped to tenth place in the Artificial Analysis Index with just 50 points. For comparison: that's nine points less than GPT-5.6 Sol and a full ten less than Claude Fable5. This isn't the first time Google has stumbled in the AI race. The company may have enormous infrastructure and a massive user base (Gemini has over 750 million users), but in recent months it has repeatedly lagged behind in the pace of innovation. While OpenAI and Anthropic churn out new models every few weeks, Google seems to have lost its direction.

What this means for the market — and for you

The reshaping of the AI model landscape has several practical implications that Czech users and businesses will also feel: 1. AI service prices will go down. When Meta offers an API at one-tenth the price of Anthropic and SpaceXAI at similarly low rates, competitive pressure forces even the pricier players to cut prices. GPT-5.6 Sol may not be the cheapest, but it already offers a better price/performance ratio than Fable5. 2. A two-tier market is emerging. Companies will choose based on need: for demanding tasks (programming, complex analyses) they'll deploy Fable5 or Sol, while for routine automation and text processing, Muse Spark or Grok will suffice. It's a similar principle to cloud services — you only pay for the performance you actually need. 3. Czech is available across all models, but quality varies. GPT-5.6 and Claude Fable5 handle Czech at a very good level — translations, summarization, and conversations in Czech are fluent and natural. Muse Spark and Grok 4.5 also support Czech, but the quality of language output lags behind the leaders. For Czech companies looking to integrate AI into customer support or internal tools, this is a key differentiator. 4. European regulation is keeping pace for now. All mentioned models are available in the EU, including the Czech Republic. Under current EU AI Act rules, they are classified as General Purpose AI models and are subject to standard transparency requirements. Additionally, since June, the US government has required developers to undergo a safety review before releasing frontier models — a system that closely resembles the European approach.

Comparison table: how the models stack up

Model	Terminal-Bench 2.1	AI Index	Agent Final Exam	API (USD / 1M tokens)
GPT-5.6 Sol (OpenAI)	88.8%	59	52.7%	$1,087 / test
Claude Fable5 (Anthropic)	88.0%	60	40.5%	10/50 input/output
Muse Spark 1.1 (Meta)	69.2%	—	—	1.25–4.25
Grok 4.5 (SpaceXAI)	—	54	—	2–6
Gemini 3.5 Flash (Google)	—	50	—	—

Who will win — performance, or price?

The answer isn't straightforward. For large enterprises and demanding developers, performance will remain decisive — and there, Anthropic and OpenAI lead today. For startups, smaller companies, and mass deployment, however, price will be the deciding factor, and that's where Meta and SpaceXAI excel. Interestingly, both camps actually complement each other. SpaceXAI leases compute power to Anthropic, and Meta is considering selling excess infrastructure. The market thus isn't splintering into hostile blocs, but rather naturally specializing — much like how premium brands and affordable car manufacturers established themselves side by side in the automotive industry. For the average ChatGPT, Claude, or Gemini user, nothing fundamental is changing for now — models are still available through familiar interfaces, and subscription prices remain stable. The real battle is being fought over enterprise customers and developers who integrate models into their own applications. And there, every dollar per token will matter.