On 22 September in US time, Anthropic released Claude Opus 5.5, and about ninety minutes later OpenAI launched GPT-6 Sol and GPT-6 Luna. Both made low price the selling point. GPT-6 Sol cut input and output tokens per million to 2 dollars and 10 dollars, half the prior GPT-5.6 Sol, and OpenAI called it a long-term price, not a promotion. Luna’s input fell from 0.2 to 0.1 dollars and output from 1.2 to 0.5, a cut up to 58 per cent. Opus 5.5 dropped to 4 dollars input and 20 dollars output, 20 per cent below Opus 5.
Cheap models are also taking the work of pricey ones. By the firms’ own benchmarks Sol matches higher-tier models on some tasks at lower cost, and Opus 5.5 beats Fable 5.1 on several tests. OpenAI improved caching to cut repeated work, and Opus 5.5 uses fewer tokens, so a typical task costs about 40 per cent less than Opus 5. For the same budget, users do more.
The expensive model is no longer the default
The new models let users buy yesterday’s premium power for less. Anthropic’s Opus 5.5 lifts coding, agent use, computer use and maths, and on nine Anthropic benchmarks beats Fable 5.1, released three weeks earlier, while priced at only 40 per cent of it. Anthropic warns the real gap is smaller than the scores suggest.
OpenAI’s GPT-6 Sol and Luna push flagship skill down to cheaper tiers, using the training method of GPT-6 Astra. Sol halves factual errors versus the last generation and nears dearer models on coding and computer use. The reason they are not priced higher is the product line itself. Fable is Anthropic’s dearest series for long hard tasks, yet Ramp data shows that a month after Fable 5 launched it was only 6 per cent of the tokens businesses bought from Anthropic. Most work does not need the top tier.
OpenAI splits by tier, with Astra keeping the top slot and Sol and Luna taking routine work. The dearest model then appears only at a few steps. Sol and Luna chase the high-volume, repeating calls that are price sensitive, so lower cost is what gets firms to embed them in daily flows.
Buyers are also sharper. Ramp’s September report shows the effective price per million tokens fell from 1.15 dollars in March to 0.68 dollars, down 41 per cent in half a year, with standard models like GPT-5.6 Terra and Claude Sonnet driving volume. The share of tokens from frontier models like Opus, Fable and Sol slipped from an August peak of 53 per cent to 45 per cent, and some firms now set standard models as the default.
How long the lead can hold
Price pulls users in, but keeping them depends on whether the model actually saves time and money in real work, and on rising competition. On the open side, Moonshot’s Kimi K3 scored 57 on the Artificial Analysis intelligence index, third overall and first among open models, and topped the Arena frontend-coding list above Fable 5 and GPT-5.6 Sol, the first open model to do so. Open weights let firms run models on their own servers with data staying in-house and no single vendor lock, with clearer pricing.
On the closed side, Grok and Muse Spark are winning tries. After SpaceX bought Cursor, Grok keeps a fast update cadence and leans on Cursor’s developer and code data. Grok 4.7 lifted nine points on Artificial Analysis coding-agent tests, past GPT-5.6 Sol. Meta narrowed the gap with Muse Spark 1.3, its fourth update in five months, matching GPT-5.6 Sol on the index at lower average cost, a sign it has joined the frontier.
OpenAI and Anthropic still hold two cards, an installed enterprise base and delivery quality on hard tasks. Ramp’s September data shows only 6.4 per cent of paying AI firms use open platforms, while Anthropic and OpenAI sit near 40 per cent adoption each. Lower price helps defend that base, but the final stay-or-go rests on results. On hard jobs like handling large files or editing complex code, the premium model can still charge if it is steadier and faster.
Two kinds of rival can take share. One is a tech giant with the enterprise door, such as Google, whose Gemini Enterprise connects Workspace, Microsoft 365 and Jira. The other is open-source labs including DeepSeek, Alibaba, Zhipu and Moonshot. Cloud vendors now close the deploy-and-trust gap for them. On 18 September Amazon Web Services added Kimi K3 to Bedrock, so firms call it through AWS with the same access control, encryption and audit as Claude, putting open models against Claude on equal terms.
For two firms edging toward public listing, the answer shows up in the valuation the market grants. Price can narrow the gap, but holding enterprise budgets takes steady wins on the work clients care about. The cheaper tier keeps daily calls, the premium model keeps hard tasks through quality, and the open and cloud camps keep pressing from outside.
Editor’s note: This is an adapted translation of the original AIX Finance report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1080120084_120829667.