{"slug": "why-are-opus-5-5-and-gpt-6-sol-cheaper", "title": "Why are Opus 5.5 and GPT-6 Sol cheaper?", "summary": "Anthropic's Opus 5.5 and OpenAI's GPT-6 Sol both cut API prices and cost roughly half as much per coding task as the models they replace, according to a 12-task benchmark run three times each on all four models. Opus 5.5 dropped from $5 to $4 per million input tokens and $25 to $20 per million output tokens (20% cheaper, 60% cheaper on cached input at $0.20), while GPT-6 Sol fell 50% across the board to $2 input and $10 output, and the new models wrote 36% and 12% fewer output tokens respectively. Artificial Analysis's larger test found Opus 5.5 costs 0.60x Opus 5 at medium effort and 0.50x at high effort, but 1.20x at max effort, where Simon Willison reported it exhausted its 128,000-token thinking budget on a simple drawing prompt without answering.", "body_md": "Anthropic released Opus 5.5 and OpenAI released GPT-6 Sol yesterday. Both cost less per token than the models they replace.\n\n| Model | Input | Output | Cached input | \n|---|---|---|---|\n| Opus 5 | $5 | $25 | $0.50 | \n| Opus 5.5 | $4 | $20 | $0.20 | \n| GPT-5.6 Sol | $4 | $20 |  | \n| GPT-6 Sol | $2 | $10 |  | \n\nPrices per million tokens, from [Anthropic's pricing page](https://platform.claude.com/docs/en/about-claude/pricing) and [OpenAI's launch coverage](https://venturebeat.com/technology/openai-releases-gpt-6-sol-and-luna-models-slashing-api-costs-50-or-more). Opus 5.5 is 20% cheaper on input and output and 60% cheaper on cached input. GPT-6 Sol is 50% cheaper across the board.\n\nI wanted to know two things: do you actually pay less for the same work, and where do the savings come from?\n\n## Do you pay less?\n\nA cheaper price per token doesn't mean a cheaper model. A new model can write more tokens. Price per task is a better measurement of how efficient a model is.\n\nI ran the same 12 coding tasks, three times each, on all four models: bug fixes, small features, refactors, tests, and docs. Opus ran in Claude Code and Sol ran in Codex, each at its default effort.\n\n| Model | Cost per task | Output tokens per task | \n|---|---|---|\n| Opus 5 | $0.173 | 2,227 | \n| Opus 5.5 | $0.088 | 1,434 | \n| GPT-5.6 Sol | $0.225 | 2,521 | \n| GPT-6 Sol | $0.118 | 2,178 | \n\nBoth new models cost about half as much per task.\n\nFor GPT-6 Sol, most of the savings came from the price cut. It also wrote 12% fewer tokens compared to GPT-5.6\n\nOpus 5.5 on average wrote 36% fewer tokens. So the efficiency was a combination of the 20% price cut and the model writing fewer tokens.\n\n[Artificial Analysis](https://artificialanalysis.ai/models/claude-opus-5-5) found the same thing on its much larger test:\n\n| Effort | Opus 5 | Opus 5.5 | Opus 5.5 vs Opus 5 | \n|---|---|---|---|\n| Medium | $2,732 | $1,627 | 0.60x | \n| High | $4,332 | $2,172 | 0.50x | \n| Max | $7,275 | $8,708 | 1.20x | \n\nAt max effort, Opus 5.5 costs more. It thinks much longer. Simon Willison [reported](https://news.ycombinator.com/item?id=49804316) that it used up its whole 128,000-token thinking budget on a simple drawing prompt without answering. \n\n## What makes a model cheaper\n\nThe cost of a task is the price per token times the number of tokens the model writes. To cut it, a new model can lower the price per token, write fewer tokens, or both.\n\nWriting fewer tokens comes from how the model is trained. We can measure it, but we can't see inside the model.\n\nThe lab sets the price per token. Under it is the cost of making a token: an hour of GPU time divided by the tokens the GPU makes in that hour. That cost we can test.\n\nI tested four changes on open Qwen models, on a rented GPU serving many users at once:\n\n| Change | How much cheaper to make each token | The catch | \n|---|---|---|\n| 1. Use fewer weights for each token | 2x | The whole file still has to fit in memory | \n| 2. Keep less state per token | 2.4x on long conversations | Some accuracy loss on long contexts | \n| 3. Use a newer GPU (B200 instead of H100) | 12% | Costs 1.58x more per hour | \n| 4. Use a newer GPU plus smaller weights | 23% | Some accuracy loss, and it only works on the newest chips | \n\nThe first two changes to the model saved far more than the new chip. DeepSeek, Kimi, and Qwen made each new generation cheaper mostly by redesigning the model. They all now use 3 to 5% of their weights per token, and DeepSeek cut its saved state by 93%.\n\n**Fewer weights for each token.** The model holds many sets of weights and uses only a few for each token.\n\n**Less state per token.** For each token in a conversation, the model keeps some numbers to look back at. Store them in half the space and twice as many conversations fit on one GPU.\n\n**Newer GPU.** A B200 costs 1.58x an H100 per hour but made 1.81x the tokens, so each token is cheaper.\n\n**Smaller weights.** Storing each weight in less space means less to move for each token. It only pays off on chips built for it. On the H100, it made tokens more expensive.\n\n## What Opus 5.5 and GPT-6 Sol might be doing\n\nWe don’t know for sure what improvements these models made. Here is what I could find from their announcements:\n\n**Opus 5.5:** Anthropic says it \"requires less compute to serve than Opus 5\" and writes output \"more than 30% faster.\" That fits the first change, using fewer weights for each token.\n\nAnthropic also cut the cached input price 60%. That fits the second change. Less state per token makes cached tokens cheaper to store and read back.\n\nA newer chip could be part of it too. And the model writes fewer tokens per task, which comes from training.\n\n**GPT-6 Sol.** OpenAI credits \"improvements to inference and caching.\" It says more of each prompt now comes from the cache instead of being processed fresh, which cuts the cost of input. It doesn't say what changed in the model.\n\nWhat the tests do show is how much each change is worth. Changing the model can halve the cost of making a token. A newer chip alone can't.", "url": "https://wpnews.pro/news/why-are-opus-5-5-and-gpt-6-sol-cheaper", "canonical_source": "https://www.claudecodecamp.com/p/why-new-models-get-cheaper", "published_at": "2026-09-23 15:15:34+00:00", "updated_at": "2026-09-23 15:30:30.223571+00:00", "lang": "en", "topics": ["large-language-models", "ai-products", "ai-infrastructure", "ai-chips", "ai-research"], "entities": ["Anthropic", "Opus 5.5", "OpenAI", "GPT-6 Sol", "Opus 5", "GPT-5.6 Sol", "Artificial Analysis", "Simon Willison"], "alternates": {"html": "https://wpnews.pro/news/why-are-opus-5-5-and-gpt-6-sol-cheaper", "markdown": "https://wpnews.pro/news/why-are-opus-5-5-and-gpt-6-sol-cheaper.md", "text": "https://wpnews.pro/news/why-are-opus-5-5-and-gpt-6-sol-cheaper.txt", "jsonld": "https://wpnews.pro/news/why-are-opus-5-5-and-gpt-6-sol-cheaper.jsonld"}}