Convly Original Study · Updated 2026
We cross-referenced live API pricing for 30 leading AI models with the independent Artificial Analysis Intelligence Index to answer one question — which model gives you the most intelligence per dollar? The answer overturns “you get what you pay for.”
AI model pricing in 2026 spans a staggering range — yet the most expensive model is nowhere near 100× smarter than the cheapest. We combined the live pricing in our AI models database with the independent Artificial Analysis Intelligence Index to answer one question: which model gives you the most intelligence per dollar? The answer overturns the assumption that you get what you pay for.
Key takeaways
- 114× price spread. On a blended-cost basis, frontier API models range from ~$0.18 to ~$20 per million tokens.
- Value ≠ price. The cheapest capable models deliver 30–40× more intelligence per dollar than the priciest frontier models.
- DeepSeek V4-Flash returns roughly 67% of Claude Opus 4.8’s measured intelligence at ~1.8% of the cost.
- Paying top dollar buys the last few points of capability — not proportional intelligence.
- For most workloads, a mid-tier or open model is the rational default; reserve frontier models for the hardest tasks.
Update — July 2026: an open model now tops the table
Since this study went live, Moonshot released Kimi K3 (16 July 2026) — a 2.8-trillion-parameter open-weight mixture-of-experts scoring 57 on the Artificial Analysis Intelligence Index, edging past Claude Opus 4.8 (55.7). It is the first open-weight model to outscore a Western frontier flagship, and at $3/$15 per 1M tokens it returns roughly 1.7× the intelligence per dollar of Opus 4.8.
It does not move the headline spread below, though: K3 lands mid-table on value at 6.3 intelligence points per blended dollar. GLM 5.2 still returns about 2.8× more capability per dollar, and DeepSeek V4-Flash roughly 30× more. The ultra-cheap-Chinese-model era looks to be ending too — K3 costs about 3× its own predecessor. Full analysis: Kimi K3 explained · K3 vs Opus 4.8.
How we measured it
- Pricing — live input/output API prices (per 1M tokens) from the Convly AI models database.
- Blended cost — a 3:1 input-to-output token ratio (typical of real API traffic):
blended = (3 × input + output) / 4. This makes models with cheap input but pricey output directly comparable. - Performance — the independent Artificial Analysis Intelligence Index (a composite across reasoning, coding, math and knowledge), used only where a current score is published for the exact model.
- Value — Intelligence Index ÷ blended cost.
The 2026 blended-cost index
| Model | Input $/1M | Output $/1M | Blended $/1M |
|---|---|---|---|
| DeepSeek V4-Flash | $0.14 | $0.28 | $0.18 |
| DeepSeek V4-Pro | $0.435 | $0.87 | $0.54 |
| Claude Haiku 4.5 | $1.00 | $5.00 | $2.00 |
| Gemini 3.5 Flash | $1.50 | $9.00 | $3.38 |
| Gemini 3.1 Pro | $2.00 | $12.00 | $4.50 |
| Claude Sonnet 4.6 | $3.00 | $15.00 | $6.00 |
| Claude Opus 4.8 | $5.00 | $25.00 | $10.00 |
| GPT-5.5 | $5.00 | $30.00 | $11.25 |
| Claude Fable 5 | $10.00 | $50.00 | $20.00 |
The spread is the story: Claude Fable 5 costs 114× more per blended token than DeepSeek V4-Flash. Want your own numbers? Try our AI API cost calculator.
Intelligence per dollar — the value ranking
Now we divide capability by cost. Using the Artificial Analysis Intelligence Index (higher = more capable) against blended cost, for the models with a published current score:
| Model | Intelligence Index | Blended $/1M | Intelligence per $ |
|---|---|---|---|
| DeepSeek V4-Flash | 37.4 | $0.18 | ≈ 208 |
| Claude Opus 4.8 | 55.7 | $10.00 | ≈ 5.6 |
| GPT-5.5 | 54.8 | $11.25 | ≈ 4.9 |
The finding is stark: DeepSeek V4-Flash delivers roughly 37× more intelligence per dollar than Claude Opus 4.8 — while still scoring about two-thirds of Opus’s raw intelligence. The frontier premium is real, but in capability terms it is small; in price terms it is enormous.
What this means for you
- The frontier premium buys the last points, not proportional intelligence. You pay 30–40× more for roughly the top third of capability.
- For high-volume production — chat, classification, extraction, RAG — a cheap or open model is the rational default.
- Reserve frontier models (Opus 4.8, GPT-5.5, Fable 5) for the hardest reasoning and long-horizon agentic work, where the last points genuinely matter.
- Self-hosting open weights changes the math entirely — your cost becomes hardware and electricity, not per-token. See our VRAM calculator to check what you can run locally.
Caveats
The Intelligence Index is one composite proxy — your task may weight coding, long context or multimodality differently, and frontier models lead by more on the hardest agentic tasks. Prices also change, and caching or batch discounts can cut real bills by 50–90%. Treat this as a directional value map, not a verdict for every workload.
Bottom line
In 2026, “you get what you pay for” is simply false for AI APIs. The cheapest capable models deliver the vast majority of frontier intelligence at a tiny fraction of the cost. Match the model to the task — and run your own usage through our cost calculator before you commit.
Sources: Artificial Analysis Intelligence Index (intelligence scores); Convly AI models database (pricing). Figures current as of June 2026.
Download the data — free & open (CC-BY-4.0)
The full 30-model dataset (prices, specs, intelligence scores and value ratios) is open for anyone to analyse or republish with attribution:
Cite or republish this study
Free to cite, quote or embed with a link back. Suggested credit:
“The AI Price/Performance Index 2026,” Convly — https://convly.ai/ai-price-performance-index-2026/
Headline finding for reference: a 114× blended-cost spread separates the cheapest capable AI model ($0.18/1M tokens) from the most expensive ($20/1M), yet the best-value model returns roughly 37× the intelligence-per-dollar of the priciest frontier model while still scoring about two-thirds of its raw intelligence. Journalists and researchers: the underlying data is downloadable above, and we’re happy to share the full methodology or a high-resolution chart on request.
Get the numbers before everyone else
One email a week: which AI models changed price, what the new benchmarks actually measured, and which GPU is worth buying. No hype, no filler.
Free. Unsubscribe in one click. We never sell or share your address.
