What coding agents cost on a subscription

Subscription assumptions compiled on 4 September 2026; benchmark data and local usage audited on 6 September 2026 from community usage reports and independent experiments. These are modeled subscription economics, not lab-published token entitlements.

How the estimate works

Estimated subscription cost per task = benchmark API cost per task × monthly subscription price ÷ estimated monthly API-equivalent token value.

This preserves the benchmark's actual token-cost mix in the numerator, rather than pretending cached input, fresh input and output tokens are interchangeable. A dollar of API-equivalent usage means token volume valued at API retail rates, not the provider's compute cost or spendable API credit.

The chart allocates the entire plan price to coding and assumes the estimated monthly usage is consumed. Using half of that amount doubles the effective cost. Each model is an alternative use of the same subscription; allowances must not be added together. A task is a benchmark attempt, including unsuccessful attempts, not a guaranteed successful fix.

Four Astra settings (low, medium, high and xhigh) retain approximate scores and displayed API costs from the supplied screenshot; the other seventeen configurations use exact values fetched on 6 September. In the original screenshot, Kimi ($3.08) and Fable 5 ($11.68) are corroborated by the Artificial Analysis dataset. The screenshot's duplicated Astra “high” label is resolved as high ($2.89) and xhigh ($3.27), following its lower chart. Opus $8.17 is xhigh in the lower chart and current dataset. These corrections preserve the plotted configurations. This is an independent redraw, not an Artificial Analysis subscription benchmark.

The update adds Gemini 3.7 and 3.8 Flash using the dataset's exact Antigravity SDK rows: 56.57 and 59.09 points, respectively. These differ from the screenshot's OpenCode results because the harness matters. Seventeen configurations now use exact benchmark values fetched on 6 September 2026. The four remaining configurations retain the original screenshot snapshot, identified in the downloadable data. GLM/z.ai and Muse Code now have explicit coverage below; missing measurements remain unknown rather than being assigned invented chart coordinates.

Model assumptions

Plan / model family	Fee	Central monthly API-equivalent value	Sensitivity envelope
Codex Pro	$200	$7,000	$5,000–$10,000
Claude Max, Opus proxy	$200	$7,000	$4,000–$8,000
Claude Max, Fable proxy	$200	$7,500	$2,000–$9,500
Kimi Vivace, K3	$199	$1,050	$750–$1,500
SuperGrok, Grok Build 4.5	$30	$500	$300–$750
Antigravity Ultra, Gemini Flash	$100	$3,780	$1,890–$5,671

Central values are rounded analyst choices anchored to the reports below. The ranges are deliberately broad sensitivity scenarios, not confidence intervals, a survey distribution, or minimum/maximum contractual quotas. No community sample establishes identical economics across every model or reasoning setting. Astra uses a 5.6-family proxy; Fable 5.1 uses a Fable 5 proxy. Different cache mixtures and quota accounting can change these ratios materially.

What subscribers actually measured

Codex. August quota measurements pair token logs with quota changes. One Pro 5x user inferred about 10,037 credits/week, and a commenter reconstructed 12,160 credits in a fully exhausted window. Using the thread's $0.04 credit valuation, multiplying 5x capacity by four for 20x, and 52/12 weeks per month yields about $6,960–$8,431/month. The fourfold scaling is an assumption, not a directly measured Pro 20x quota. The $7,000 central case rounds the lower estimate. A separate recent logged month reports $4,615–$4,828 with 6.69B input tokens, of which 6.29B were cached, and 20.85M output tokens. That is actual consumption, not an exhausted-capacity measurement.

Claude / Opus. One August Max subscriber reports $7,470 API-equivalent in a month on the $200 plan, after deduplicating session logs and hitting limits repeatedly. Another usage report reports $6,790 over 30 days and $1,835 over seven days. Both involve mixed models, so $7,000 is a heavy-use proxy for Opus, not an Opus-only quota measurement. An independent earlier logged experiment extrapolates roughly $4,080/month/account across three accounts, supporting the lower scenario. The reported raw total was dominated by cached tokens.

Fable. A July Fable 5 user reports $385 API-equivalent consumed 22% of the dedicated weekly Fable allowance: $385 / 0.22 × 52/12 ≈ $7,583/month. This anchors the rounded $7,500 central case. Another July experiment reports $628–$775 for 35% of the dedicated allowance, implying roughly $7,775–$9,595/month. These already account for the dedicated bucket; halving them again would double-count its restriction. Conversely, August developer measurements found Fable consumed approximately twice the weekly percentage per API dollar as Opus in the sampled workloads, and an independent synthesis modeled only about $2,365/month. This disagreement motivates the very wide range. The central estimate is an older, favorable, cache-heavy scenario and is not a verified current Fable 5.1 capacity. Fable 5.1 was too new for a full-month user log.

Kimi. An instrumented subscriber report records 405M raw tokens across August 8–9, including 11.7M fresh-input-plus-output tokens, consuming 61% of weekly quota. A linear extrapolation gives about 664M raw/week or 2.88B/month. The source has a two-seat ledger ambiguity, so this is corroboration, not a precise single-seat entitlement. The community comparison estimates Vivace/K3 at $241–$250 API-equivalent/week, about $1,044–$1,083/month. It describes an August 13 measurement but does not link its exact originating report. $1,050 is therefore less strongly grounded than the Codex/Claude central estimates; $750–$1,500 is an analyst sensitivity envelope. Raw token equivalents depend on the cache mix, not just model name.

Grok Build. One SuperGrok subscriber's logged weekly experiment reports $115.38 API-equivalent over a full weekly allowance on the $30 plan, implying about $500/month. A separate Grok Build session log reports $20.82 in its internal product meter for 16 percentage points of weekly allowance; extrapolation gives approximately $564/month. The latter is a product-meter estimate, not necessarily API retail value, so it corroborates the rough scale only. The model mix and complete cost calculation are not independently reproducible. The central $500 estimate therefore has an analyst sensitivity envelope of $300–$750/month, giving $0.10–$0.24 per benchmark task. The chart explicitly labels the actual $30 plan instead of inventing a $200 product.

Gemini through Antigravity

Provisional estimate: $0.037/task for Gemini 3.7 Flash high; $0.053 for 3.8 Flash high. The scenario ranges are $0.025–$0.074 and $0.035–$0.106.

One subscriber's session-log experiment estimates 5B raw tokens/week on the $100 Ultra tier while using Gemini 3.7 Flash. Ultra was inferred from two nearly complete five-hour windows, not a fully exhausted week. The author reports a regional discount; the chart uses the stated US tier price. Their Gemini cache/input/output split is absent. Their 98.2% cache figure refers to Claude, so it is not used here.

Our assumption: value the reported raw capacity at the token mix measured in Artificial Analysis's native Antigravity SDK benchmark for Gemini 3.7 Flash. That benchmark averages 8.003M prompt-plus-output tokens and $1.3963 API cost per task. Therefore 5B × 52/12 ÷ 8.003M × $1.3963 gives $3,780/month API-equivalent. This is a calculated scenario, not a subscriber-reported dollar allowance. Gemini 3.8 uses the same monetary denominator with its own $2.0090 API task cost. Assuming comparable token mix across workloads and transferable quota economics across these models is uncertain.

For these Gemini records, the input counter already includes cached input. Adding the cache counter again would double-count tokens. The 3.8 record reproduces its API cost when cached input is subtracted from total input before fresh-input pricing, using the rates in the AA Gemini 3.8 profile. Its mean totalTokens field adds cache again and is not used. The calibration uses inputTokens plus outputTokens.

The 2.5–7.5B/week sensitivity band is an analyst choice around the partial-window estimate. Different token mixtures can shift costs beyond it. The actual $100 tier avoids extrapolating unmeasured $200 capacity. It is priced separately from its own $20 Pro tier; they are not stacked.

GLM through z.ai

Quota estimate available; cost per task not yet comparable. Community discussions list Lite $18, Pro $80 and Max $168. We use that source only for the plan prices, not its vendor-derived allowance figures.

A GLM 5.3 Lite subscriber reports 48M tokens consuming 44% of weekly quota, with a 95% cache hit rate. Extrapolating gives 109M raw tokens/week, or 473M/month on Lite. This is a single short self-report without raw logs. It does not establish the output share, peak-hour effects, or the capacity of Pro and Max. Multiplying the Lite measurement by advertised tier ratios would add an unverified assumption.

The checked Coding Agent Index dataset has GLM 5.2 at 43.31 points, but no GLM 5.3 row. The widely cited GLM 5.3 Intelligence Index score belongs to a different benchmark and cannot supply this chart's vertical coordinate. Assigning 5.2's task cost or score to 5.3 would also change the model/provider configuration. GLM 5.3 remains unscored. The full dataset retains historical GLM 5.1 and 5.2 agent results with current Pro quota scenarios, as described below.

Our recorded usage, audited 6 September 2026

Recorded API-equivalent usage is not a subscription allowance. The companion graph shows totals over each stated period, without extrapolating them into guaranteed monthly capacity. Different periods and account counts are not a like-for-like efficiency ranking.

• Claude, August: $19,380.10 priced subtotal. The recovered independent audit reproduces the supplied mixed-model aggregates. Fresh-input volumes and Opus 4.8 were not fully priced. The source combines two plans and paid overages; it cannot establish a Fable 5.1 included allowance.
• Codex, August: $5,327.97 priced subtotal. A fresh scan reproduces Sol at $4,954.28 and Terra at $373.69. Minor models are excluded where matching prices were unavailable.
• Astra, 1–5 September UTC: $1,462.00. Five complete days contain 1.072B input tokens, including 1.046B cached, plus 3.013M output tokens. Per-call counters are deduplicated across replayed fork history. Inherited cumulative session counters are not summed.
• GLM family, 10 August–5 September UTC: $194.67 across 617.14M raw tokens. This includes GLM 5.2, 5.3 and Turbo across two provider IDs. GLM 5.3 alone contributes $75.49 and 233.75M tokens, starting 16 August. The coding-plan provider contributes $190.74; the offpeak-idle provider $3.93. Shared subscription ownership is unproven. GLM cached tokens are a subset of input and are priced once.

The interrupted draft described two additional points as measured Fable and Astra subscription allowances. The underlying audit did not support those denominators, so those points are removed. Fable 5.1 keeps the explicitly labeled community proxy; the personal logs remain a separate consumption graph.

Current GLM plan constraints

The graph uses the Coding Agent Index throughout, not the model Intelligence Index. The full dataset retains historical Claude Code results: GLM 5.1 at 36.91 points and $4.282 API cost per attempt; GLM 5.2 at 43.31 points and $1.9094 API cost. Applying today's $80 Pro quota envelope gives $0.529 per task for 5.1 ($0.325–$0.861) and $0.236 for 5.2 ($0.145–$0.384). The central denominator is the geometric midpoint of the $398–$1,055 full-usage envelope. This is a transfer assumption, not a measured cost for a currently selectable model.

The author reports hitting GLM plan caps. The logs corroborate capped sessions, but a five-hour cap does not establish use of every weekly allowance. Personal usage remains visible separately; the historical records use the explicit full-usage scenario.

The official Coding Plan documentation routes requests for GLM 5.1 and 5.2 to 5.3. The benchmark still contains older 5.1 and 5.2 configurations, which cannot establish 5.3 performance.

Pro allows 60,000 credits weekly and 12,000 per rolling five hours. Applying its token multipliers to official API prices, full weekly utilization gives a theoretical monthly API-equivalent bound of $398–$528 at peak rates, or $795–$1,055 entirely off peak, depending on token mix and excluding MCP use. This is a contractual scenario, not a local usage measurement. The draft's $1,000–$2,500 extrapolation ignored the weekly ceiling and is removed.

Muse Code

Plans exist; a measured token allowance is still missing. The community launch discussion describes $5 Everyday, $15 High and $50 Power plans. The $50 tier would be the relevant highest native plan; an exact $200 product is not required.

In Fazt's paid Muse Code test, the $5 plan exhausted its quota after 30–40 minutes on Muse Spark 1.2 ultra, while a large multi-application task remained incomplete. No token totals or cache split were reported. Minutes of activity cannot be converted into benchmark tasks without inventing throughput. A Muse Spark 1.3 subscriber also asks how the native meter counts Contributor usage, without a measured answer.

To calculate Muse's position, a useful report needs the native plan tier, exact model, quota percentage before and after, reset period, and input/cache/output tokens or reconstructed API value. OpenCode Go anecdotes measure a different subscription. Muse remains unpriced until that evidence is available.

Why some models remain unpriced

These remain outside the priced comparison:

• Qwen3.8 Max: A token-volume report concerns an $18 plan under a 50× promotion. Extrapolating it to normal $200 usage would mislead.
• Three Cursor CLI configurations: GPT-5.5 medium, Opus 4.7 medium and Composer 2.5 have published agent scores and API task costs, but no adopted same-model subscription calibration. They remain visible on the original API panel.

Worked examples

• Astra high: $2.89 × $200 / $7,000 = $0.083/task.
• Astra max: $4.72 × $200 / $7,000 = $0.135/task.
• Fable 5.1 max: $9.18 × $200 / $7,500 = $0.245/task, conditional on the older Fable proxy.
• Kimi K3: $3.08 × $199 / $1,050 = $0.584/task.
• Grok Build / Grok 4.5 high: $2.44 × $30 / $500 = $0.146/task.
• Antigravity / Gemini 3.7 Flash high: $1.3963 × $100 / $3,780.38 = $0.037/task.
• Antigravity / Gemini 3.8 Flash high: $2.0090 × $100 / $3,780.38 = $0.053/task.

Subscription-cost scenarios exclude tax, overage purchases, hardware, review time and other subscription benefits. The separate personal usage totals can include paid overages.

The original Artificial Analysis methodology explains its benchmark attempts and API cost accounting. Its score and task-cost measurements are the starting point; the subscription conversion is my estimate.

## Side-by-side comparison and our Cursor Pro plan

The original panel retains Artificial Analysis’s published API cost per attempt. The adjusted panel changes the cost coordinate and keeps the same score. Both panels use linear cost scales: $1–$12 for API pricing and $0.03–$0.35 for subscriptions. The lower subscription bound retains Gemini 3.7 at about $0.037. Their score axes remain identical (55–72). The frontier view shows 14 configurations on each side; the complete dataset retains 21 original configurations and 18 subscription estimates. Four Astra settings retain the original screenshot values; 17 configurations match current AA agent records. Every visible dot names its harness and model. Kimi lies beyond the subscription axis and is marked with an arrow labeled with its actual $0.584 central cost. The API panel retains its Kimi dot. Luna max, Composer and historical GLM are excluded from the view; older Cursor configurations fall below its score range. All source records remain in the full values table. Colors identify labs: OpenAI black, Anthropic orange, Gemini multicolor, Grok green and Kimi blue. Soft areas group Astra medium through max plus Sol max, all Anthropic configurations and both Geminis; they are family groupings, not statistical uncertainty. Horizontal lines show sensitivity ranges. An arrow marks a range extending beyond $0.35; the complete value remains in the point tooltip.

Our authenticated Cursor billing page confirms a paid $20 Pro invoice on 10 August 2026, covering 10 August–10 September 2026. On 6 September the Cursor Models pool was 100% used, Other Models 88.8% used, and on-demand spending disabled. The 827-row export totals 827.12M Included tokens and 37.89M Free tokens; all rows passed an additive input/cache/output check. The two pools belong to the same subscription and are not additive dollar allowances. This observation describes the current paid cycle, without assuming renewal.

Composer 2.5 Fast is a conditional same-model quota estimate. Five recorded Included calls used 506,495 fresh input tokens, 3,932,352 cache-read tokens and 44,742 output tokens. At current Cursor prices of $3, $0.50 and $15 per million respectively, their API-equivalent value is $4.156791. The billing dashboard attributes 0.9% of the Cursor Models pool to that model. Assuming linear dollar-equivalent quota use, this implies $4.156791 / 0.009 = $461.8657 of same-model API-equivalent capacity. It is not a published contractual dollar allowance.

The matched Cursor CLI – Composer 2.5 Fast agent score is 38.30075. Its original AA cost remains $0.557344 on the left panel. For the right panel we estimate $1.522275 from its pooled mean token counters at current Cursor rates: (2,153,033.9826 input − 2,095,607.8373 cached) × $3/M + cached × $0.50/M + 20,146.1462 output × $15/M. This assumes the input counter includes cache, and that the pooled means cover comparable attempts. AA aggregates metrics with missing telemetry separately, so this is approximate repricing, not a reconstruction of an exact benchmark bill.

The resulting full-fee scenario is $20 × $1.522275 / $461.8657 = $0.065918 per task. Interpreting the displayed 0.9% as a rounded 0.85–0.95% gives $0.062256–$0.069581. That range covers rounding only, not uncertainty from five calls, cache mixture, telemetry coverage, nonlinear credits or future quotas. Each full-fee model scenario allocates the same entire $20 subscription to one alternative; the pools must not be charged or summed twice.

Other Cursor agent rows remain unpriced: no separate matched Composer 2.5 quota sample, and no adopted current-rate calibration for the historical GPT-5.5 medium and Opus 4.7 medium configurations. Most recorded Cursor tokens went to Grok variants; the checked agent dataset contains no matched Cursor CLI Grok result. We do not borrow a Grok Build, Codex or Claude Code score for Cursor.

GLM 5.3 recheck

A fresh no-cache check on 6 September 2026 found 68 coding-agent records. The GLM records are Claude Code with GLM 5.1 and 5.2. GLM 5.3 occurs in the page’s generic model-catalog metadata and has a separate model profile, but no matched agent result or Coding Agent Index score. Our 233.75M recorded GLM 5.3 tokens are real usage evidence, not an agent benchmark.