Deeptech and AI · Research note · 25 Sep 2026
Two prices of thought: the frontier and the trailing curve
On 22 September Anthropic released a frontier model priced 20 percent below its predecessor. The same day Epoch AI published data showing the cost of a fixed level of AI performance falling about 47 percent a quarter. The two numbers describe different curves. Which one a business sits on decides its margin, and India's $221 billion software export sector sits on the wrong side of the faster one.
Sent. Check your inbox.
If it is not there in a minute, look in Promotions or Updates and mark it as not spam so the next one lands in the inbox.
Every number, and where it comes from
The first five figures in the note.
The full table has 8 rows. Each figure is as published on the date shown, with its denominator.
| Figure | What it measures | Source |
|---|---|---|
| $4 / $20 | Claude Opus 5.5 list price per million input and output tokens from 22 Sep 2026, 20% below Opus 5's $5 / $25. Cache reads $0.20 per million, 60% lower. | Anthropic, Introducing Claude Opus 5.5, 22 Sep 2026 |
| 40% | Anthropic's own test result for how much less Opus 5.5 costs than Opus 5 on typical workloads at default settings, attributed to lower per-token price plus fewer tokens per task. Vendor measurement, not a list price. | Anthropic, 22 Sep 2026 |
| $30 / $60 | GPT-4 launch price per million prompt and completion tokens, March 2023. OpenAI quoted $0.03 and $0.06 per thousand tokens for the 8K-context model; the per-million figure is the equivalent. Output down two thirds and input down 87% to Opus 5.5 is author's arithmetic across two vendors. | OpenAI, GPT-4 research announcement, 14 Mar 2023 |
| $30; $25 to $20; $12 | Output list prices per million tokens on 23 Sep 2026: OpenAI GPT-5.6 Sol $30 (input $5); Anthropic Opus 5 $25 to Opus 5.5 $20; Google Gemini 3.1 Pro $12 for prompts up to 200K tokens ($18 above). Pricing pages carry no publication date; the date is the fetch date. | OpenAI API pricing; Google Gemini API pricing; Anthropic Opus 5.5 announcement and Claude Platform Docs, fetched 23 Sep 2026 |
| 47% a quarter; 13x a year | Fall in the cost of achieving a given level of AI performance since 2023, averaged across five benchmarks, measured as the cheapest model reaching each score. 66% a quarter for performance that has just debuted as state of the art; 32% a quarter for performance first reached two years earlier. | Epoch AI, Emberson and Roodman, The plunging price of thought, 22 Sep 2026 |
3 more figures in the PDF, each with its source.
Sources
9 primary releases.
- Anthropic, Introducing Claude Opus 5.5 (22 Sep 2026)Input and output tokens $4 and $20 per million, 20% less than Opus 5; cache reads $0.20, 60% less; costs 40% less to run on typical workloads; Opus 5 table row $5 / $25.anthropic.com
- Anthropic, Claude Opus 5.5, Claude Platform Docs model overview (fetched 23 Sep 2026)Context window 1M tokens; max output 128K; input $4, output $20, cache read $0.20 per million.platform.claude.com
- OpenAI, GPT-4 (14 Mar 2023)Pricing is $0.03 per 1k prompt tokens and $0.06 per 1k completion tokens.openai.com
- OpenAI, API pricing (fetched 23 Sep 2026)GPT-5.6 Sol: input $5.00, cached input $0.50, output $30.00 per million tokens; standard rates for context under 270K.openai.com
- Google, Gemini API pricing (fetched 23 Sep 2026)gemini-3.1-pro-preview: input $2.00 and output $12.00 per million tokens for prompts up to 200K; $4.00 and $18.00 above.ai.google.dev
- Epoch AI, The plunging price of thought (22 Sep 2026)Cost of a given level of AI performance down about 47% per quarter since 2023, 13x per year; 66% per quarter for newly state-of-the-art performance, 32% two years later; o3 to GPT-5.6 Luna 725-fold drop on GPQA Diamond.epoch.ai
- Epoch AI, LLM inference prices have fallen rapidly but unequally across tasks (12 Mar 2025)Declines of 9x to 900x per year depending on the milestone; 40x per year for GPT-4-level performance on PhD-level science questions.epoch.ai
- a16z, LLMflation: LLM inference cost is going down fast (12 Nov 2024)For an LLM of equivalent performance, the cost is decreasing by 10x every year.a16z.com
- Reserve Bank of India, Survey on Computer Software and Information Technology Enabled Services Exports: 2025-26 (18 Sep 2026)Software services exports up 8.2% to $221.4 billion, excluding sales through overseas commercial presence; United States share 54.1%.rbi.org.in