Analysis briefing · model economics

Do Claude plans really give 2–3× more value? Half-true.

The screenshot is right on the raw number: a $20 Claude Pro plan buys ≈2.9B tokens/month against ≈1.0B on ChatGPT Plus — 2.9×. But Artificial Analysis' official per-task data shows Claude Opus 5.5 (Max) burning ≈119k output tokens to finish one standard task, versus ≈38k for GPT-6.1 Sol (Max). Divide one by the other and the advantage disappears: at the same effort level, a Claude plan completes ≈0.9 tasks for every 1 ChatGPT completes per dollar, and Opus costs 8.3× more per task in API terms.

6 October 2026 SemiAnalysis Tokenomics Model (published 5 Oct 2026) · Artificial Analysis Intelligence Index v4.3.2 (retrieved 6 Oct 2026).
Flagship configs as named in the original chart: GPT-6.1 Sol (Max) vs Claude Opus 5.5 (Max, Default Fallback).

The claim

2–3× more tokens per dollar — that part is real

SemiAnalysis limit-tested every OpenAI and Anthropic consumer plan (long-horizon agentic coding until the weekly limit) and counted every billed token: fresh input, cached input, cache writes and output. Anthropic's plans hand out ~2.8× more tokens per fee dollar and ~5.5× more API-equivalent value, because Opus 5.5 tokens also cost about twice as much each.

OpenAI · GPT-6.1 Sol Anthropic · Claude Opus 5.5

$20/mo · Plus vs Pro

2.9× more tokens

Bars scaled to the largest bar on this page (28.6B tokens).

ChatGPT Plus1.0B tokens · $211 API value
1.0B
Claude Pro2.9B tokens · $1,178 API value
2.9B

$100/mo · Pro 100 vs Max 5x

2.7× more tokens

14.0B is the second-largest bar; Claude's value ratio stays ~5.4×.

ChatGPT Pro 1005.1B tokens · $1,055 API value
5.1B
Claude Max 5x14.0B tokens · $5,725 API value
14.0B

$200/mo · Pro 200 vs Max 20x

2.8× more tokens

The headline row of the original chart: 28.6B vs 10.2B tokens.

ChatGPT Pro 20010.2B tokens · $2,084 API value
10.2B
Claude Max 20x28.6B tokens · $11,726 API value
28.6B
Tokens and API-equivalent value per plan Source: SemiAnalysis Tokenomics Model, 5 Oct 2026 · agentic workload mix: 0.4% input · 96.6% cached input · 2.6% cache writes · 0.3% output. API-equivalent value at each lab's first-party API prices.
Tier ChatGPT Claude Tokens ratio API-value ratio
$20/mo 1.0B tokens · $211 value 2.9B tokens · $1,178 value 2.9× 5.6×
$100/mo 5.1B tokens · $1,055 value 14.0B tokens · $5,725 value 2.7× 5.4×
$200/mo 10.2B tokens · $2,084 value 28.6B tokens · $11,726 value 2.8× 5.6×

The other half of the ledger

Tokens per task, measured by Artificial Analysis

AA runs every model itself over the same ten evaluations (Intelligence Index v4.3.2) and publishes output tokens per task — reasoning plus answer. This is the missing number in the token chart: how much a model burns to finish the same unit of work, at the same effort setting. Opus 5.5 (Max) needs ≈119k tokens per task. GPT-6.1 Sol (Max) needs ≈38k. 3.1×.

GPT-6.1 Sol Claude Opus 5.5

Output tokens per AA Intelligence Index task · scaled to 120k

Opus burns 2.6–3.7× more per task at every effort tier

Left bar Sol, right bar Opus, per effort setting. Values are the weighted-average output tokens per task.

Lowindex 42.1 / 42.3
4.0k
10.2k
Mediumindex 47.8 / 51.2
8.1k
25.7k
Highindex 50.2 / 53.6
13.2k
35.6k
X-highindex 51.0 / 56.0
17.6k
65.7k
Maxindex 51.8 / 57.6
38.1k
119.2k
Artificial Analysis, Intelligence Index v4.3.2, retrieved 6 Oct 2026 Sources: AA comparison: GPT-6.1 Sol (Max) vs Claude Opus 5.5 (Max, Default Fallback) · Sol release page · Opus 5.5 release page. "Default fallback" = AA's default configuration for the model.
Effort Sol index Sol tokens/task Sol $/task Opus index Opus tokens/task Opus $/task Token ratio Cost ratio
Low 42.1 4.0k $0.13 42.3 10.2k $0.55 2.6× 4.2×
Medium 47.8 8.1k $0.21 51.2 25.7k $1.34 3.2× 6.2×
High 50.2 13.2k $0.32 53.6 35.6k $1.82 2.7× 5.7×
X-high 51.0 17.6k $0.39 56.0 65.7k $3.46 3.7× 8.8×
Max 51.8 38.1k $0.72 57.6 119.2k $5.98 3.1× 8.3×

Intelligence at N tokens

The same idea as a chart: quality vs. tokens spent

Each line traces one model up its effort settings: right = more tokens, up = higher Intelligence Index. Opus 5.5 reaches a higher ceiling (58 vs 52) but pays for it in tokens — and at equal quality it can be the thriftier model (next section).

40 45 50 55 60 0 25k 50k 75k 100k 125k output tokens per AA Intelligence Index task Intelligence Index ≈equal quality Sol (Max) · 52 Opus (Max) · 58

The math the chart skips

2.9× more tokens ÷ 3.1× more tokens per task = 0.93× the work

Holding the billed-token mix constant (both plans are measured on the same agentic workload shape), what matters is tokens-in divided by tokens-per-task. The token bonus and the token burn are almost the same size — so the work you can actually get done per month is roughly a wash, and noticeably worse once you price the same work in API dollars.

Token math · $20 plan

The advantage cancels itself

Tokens in planClaude ÷ ChatGPT
2.9×
Tokens per taskOpus ÷ Sol, max effort
3.1×
Net work per monthGPT = 1.00
0.93×

2.9 ÷ 3.1 = 0.93 · GPT finishes ~7% more equivalent tasks

Dollar math · $20 plan

Worse once priced at API rates

API value in planClaude ÷ ChatGPT
5.6×
Cost per taskOpus ÷ Sol, max effort
8.3×
Net work per $GPT = 1.00
0.68×

5.6 ÷ 8.3 = 0.68 · ~32% fewer dollar-equivalent tasks

Net work per plan, both ways of counting · GPT = 100 Token path: (plan tokens ratio) ÷ (AA output tokens per task ratio, max effort 119.2k ÷ 38.1k). Dollar path: (SemiAnalysis API-value ratio) ÷ (AA cost-per-task ratio, $5.98 ÷ $0.72). Billed-token mix held constant — see caveats.
Tier Tokens ratio ÷ token/task ratio Net (token-normalized) API-value ratio ÷ $/task ratio Net ($-normalized)
$20/mo 2.9× 3.1× 93 per 100 GPT tasks 5.6× 8.3× 68 per 100 GPT tasks
$100/mo 2.7× 3.1× 88 per 100 GPT tasks 5.4× 8.3× 66 per 100 GPT tasks
$200/mo 2.8× 3.1× 90 per 100 GPT tasks 5.6× 8.3× 68 per 100 GPT tasks

Where it flips

The answer depends on the effort tier you compare

Same-tier comparisons at low and high effort actually favour Claude; x-high and max favour GPT; and comparing equal intelligence rather than equal effort is the strongest Claude case. The "it's a wash" verdict is specific to flagship-max against flagship-max — the two configs named in the original chart.

Claude tasks per 100 GPT tasks, at the $20 tier Token-normalized = plan tokens ÷ AA output tokens per task. $-normalized = SemiAnalysis API value ÷ AA cost per task. Anything above 100 favours Claude; below 100 favours GPT.
Comparison Sol tokens/task Opus tokens/task Token-normalized $-normalized
Low vs low 4.0k 10.2k 114 133
Medium vs medium 8.1k 25.7k 91 89
High vs high 13.2k 35.6k 108 98
X-high vs x-high 17.6k 65.7k 78 63
Max vs max the chart's configs 38.1k 119.2k 93 68
Equal intelligence Opus Medium ≈ Sol Max (index ~51) 38.1k 25.7k 430 303

Method, limits, sources

What is measured vs. inferred

What is measured

Tokens per plan: SemiAnalysis limit-tested each plan to exhaustion, isolated each token type against the usage meter, and normalised with an agentic-workload mix of 0.4% input / 96.6% cached input / 2.6% cache writes / 0.3% output. Tokens and $ per task: Artificial Analysis ran both models on identical evaluations; cost per task uses input, cache, reasoning and answer prices.

What is inferred

Dividing plan tokens by tokens-per-task assumes the billed token mix is constant across models. If Claude's harness re-reads more context per task — SemiAnalysis measures Claude Code at ~100:1 input:output vs Codex ~80:1 — the real skew is worse for Claude, not better. If nothing else, treat the net 0.9× as a directional correction to 2.9×, not a precise forecast.

Other things to know

  • Benchmark tasks ≠ your work. SemiAnalysis' own objection to citing AA here: "we do not believe the benchmark tasks in the AA Intelligence Index are at all representative of real work people do with LLMs." Nobody has reliable production per-task data yet; AA's is the best public stand-in.
  • Snapshots move. OpenAI halved the API-equivalent value of the $200 plan in late September 2026 and cut Sol cache pricing; Anthropic raised Opus token limits with the Opus 5.5 price cut (input/output −20%, cache reads −60%). Limits and prices change silently.
  • Effort defaults differ. AA's numbers are measured at AA's settings; subscription harnesses pick their own defaults. The chart's dashboard shows a MEDIUM preset — medium-vs-medium lands at 91/89, the same "a wash" conclusion as max-vs-max (93/68).
  • What Claude's extra money does buy: at max effort Opus 5.5 leads AA's index 58 vs 52 (+6 points), and Claude plans remain far more generous at equal intelligence. The 2.8× token number is simply not the right proxy for either of those.

Sources

Takeaway

Tokens are the billing unit, not the work unit

  • The 2–3× claim is real as a token count — and as a dollar-value count (5.4–5.6×) — but not as a work count. At the same effort setting, Opus 5.5 burns ≈3.1× more output tokens per standardized task, which all but exactly cancels the ≈2.8× token bonus.
  • Flagship-max verdict: ~0.88–0.93 AA-style tasks per Claude plan dollar versus ChatGPT (token-normalized), and 0.66–0.68 when the same work is priced at API rates. Cost per task: $5.98 vs $0.72.
  • The chart is not wrong, it is incomplete. Raw tokens per plan ignore (a) price per token, (b) tokens per task, and (c) which model you are actually buying. Add the first and Claude looks 5–6× better; add the second and it looks ~10% worse.
  • Real Claude edges: the strongest max-effort model (index 58 vs 52), and — at equal intelligence or low effort — genuinely more work per plan dollar. Real Claude costs: 8.3× per task at max effort, and a token bill that eats its own headline.

Bottom line: for the two configs in the chart, the "2–3× more value" reads ≈0.7–0.9× of ChatGPT's work per dollar. At matched quality, it reads 3–4× in Claude's favour.