Near-Frontier AI For A Fifth Of The Price: GPT-6.1 Sol And The 2026 Price War - What It Means For Your AI Budget

1/5 the price - GPT-6.1 Sol: 'near-Astra performance at one-fifth the price' - $2/M input, $10/M output tokens  ·  95% off - Cached input drops to $0.10 per million tokens - transformative for repeated workloads  ·  1M+ context - A 1,050,000-token context window with up to 128,000 output tokens  ·  8x faster, 6x price - The new Ultrafast tier - speed is now sold as a separate premium product

If you want one product launch that sums up the economics of AI in 2026, it is GPT-6.1 Sol. OpenAI launched it on 1 October with a pitch that could have been written for a finance director: near-Astra performance at one-fifth of the price. Sol costs $2 per million input tokens and $10 per million output tokens. Input that has been seen before - cached - drops to just $0.10 per million, a 95% reduction. It accepts a context window of more than a million tokens and can produce up to 128,000 tokens of output. On the same day, Anthropic released Claude Sonnet 5.5, its own capable, cost-efficient mid-tier model. And at the same event OpenAI introduced Ultrafast, a tier that generates tokens up to eight times faster than standard - priced at six times the standard rate. Taken together, these launches describe a market that has split in two. As an AI Agency London that manages AI budgets for clients, we think understanding that split is one of the most useful things a business can do this autumn, and this analysis explains it.

Why The Second-Best Model Is Now The Right Default

For two years, many businesses defaulted to the most capable model available for everything, on the reasonable logic that the best model would give the best results. That logic no longer holds. When a model offers near-frontier performance at a fifth of the price, the frontier model only earns its premium on the small share of tasks where that last margin of quality genuinely matters - complex reasoning, high-stakes analysis, difficult judgement. For everything else - drafting, summarising, classifying, extracting data from documents, answering questions from your knowledge base, routine analysis - the near-frontier model gives results that are indistinguishable in practice, at a fraction of the cost. The financial effect of switching your default is large: if most of your AI usage moves to a model costing a fifth as much, your AI bill can fall dramatically with no visible change in quality. The right approach is to make the capable, cheaper model your default and to reserve the frontier model, deliberately, for the specific tasks that have been shown to need it.

• Default to the near-frontier tier - drafting, summarising, extraction, classification and Q&A rarely need the most expensive model.
• Reserve the frontier for proven need - complex reasoning and high-stakes analysis, identified by testing rather than assumption.
• Cache what repeats - long instructions, reference documents and knowledge bases reused across requests cost 95% less when cached.
• Buy speed only where it pays - Ultrafast is worth six times the price only where seconds genuinely matter to customers.
• Re-test regularly - prices and models move monthly. Yesterday's right choice may be today's overspend.

Caching And Speed: The Two Quiet Changes

Two details of these launches deserve more attention than the headline price. The first is caching. Many business AI workloads send the same large block of material with every request - a long set of instructions, a policy manual, a product catalogue, a knowledge base. With cached input at $0.10 per million tokens, a 95% reduction, that repeated material becomes almost free after the first time. For customer service, document processing and internal assistants built on a fixed body of knowledge, the effect on cost can be bigger than the headline price cut, but only if the system is designed to take advantage of it - with stable content placed where it can be cached. The second change is speed. By pricing Ultrafast at six times the standard rate, OpenAI has made speed a separate product with its own price. That is useful clarity: most background work - overnight processing, reports, batch document handling - gains nothing from extra speed, while a live voice agent or a customer-facing assistant may genuinely convert better when it responds instantly. Businesses should buy speed deliberately for the interactions where customers feel it, and nowhere else.

The Bottom Line

GPT-6.1 Sol - near-frontier performance at a fifth of the price, with cached input 95% cheaper, a million-token context and 128,000-token outputs - launched the same day as Claude Sonnet 5.5 and alongside an Ultrafast tier selling eight times the speed at six times the price, and together they describe a two-tier market: capable AI getting cheaper fast, the very best and fastest holding their premium. For UK businesses the practical consequences are clear. Make the capable, cheaper model your default and reserve the frontier for tasks proven to need it. Design systems so repeated material is cached. Buy speed only where customers feel it, and keep the flexibility to switch models as prices move monthly. Businesses that do this can cut AI costs dramatically without losing quality, and those that do not will keep paying frontier prices for routine work. Designing AI systems that capture these savings is exactly what we do.

References & Further Reading

• Emergent - OpenAI DevDay 2026: every announcement (GPT-6.1 Sol pricing, caching, context, Ultrafast, Claude Sonnet 5.5 same day): https://emergent.sh/news/openai-devday-2026
• Pulse 2.0 - OpenAI unveils Dots, GPT-6.1 Sol, $500 Pro tier and new enterprise AI tools: https://pulse2.com/openai-unveils-dots-gpt-6-1-sol-500-pro-tier-and-new-enterprise-ai-tools/
• Latent Space - AINews: OpenAI DevDay 2026: https://www.latent.space/p/ainews-openai-devday-2026-dots-61
• Digital Applied - AI model releases: October 2026 tracker and dated ledger: https://www.digitalapplied.com/blog/ai-model-releases-october-2026-tracker
• Blog.mean.ceo - OpenAI news, October 2026: https://blog.mean.ceo/open-ai-news-october-2026/