Flat subscriptions can undercut pay-per-token pricing once you clear break-even — this ranks how, with the math shown.
Token Perks tracks 129 ways to buy frontier-model tokens — the small units AI providers bill by — across subscriptions, pay-as-you-go API, credit systems, coding-tool plans, and free tiers, and ranks what can be ranked on effective cost. Intelligence scores are quoted from Artificial Analysis with a citation on every number; our own weighted ranking is documented but deliberately withheld. The full ledger lives on the universe page. Verified Sep 6–7 2026.
Every per-task figure below uses one illustrative reference rate: $0.80 per finished task — about 100,000 tokens at the $8-per-million reference blend; your real tasks will cost more or less.
Not sure which is you: buying for someone else or not deep in APIs — start with the 9 tracked consumer offers and the gift guide; building on AI APIs — the leaderboard ranks every pay-per-token route, and providers list the same data per company; optimizing batch jobs — the Batch $/M column shows discounted arithmetic only where the provider publishes a modifier (a dash means none was published, not zero).
Cheapest free route
$0.0000
Muse Spark 1.3 Contributor Free$0 metering while the promo lasts · read 2026-09-06
Cheapest paid, per million tokens
$0.119 /M
GLM-5.3-Flashblended (3×in+out)/4 · promo-priced; read 2026-09-07
Cheapest paid, per finished task
$0.012
Z.ai GLM-5.3-Flash100k-token reference task — your mix will differ · Sep 6–7 2026
Best per task, overall right now
$0.0000
Muse Spark 1.3 Contributor Free (free promo)metered $0 while it lasts; paid floor $0.012/task
Every figure above is derived from the dated ledger at render time — per-task values use the 100k-token reference. Re-verify at official terms before paying.
9 offers tracked · sorted by estimated effective cost
| Offer | Price | Annual effective | Est. $ per 100 tasks | Key limit | Renewal behavior | Verified |
|---|---|---|---|---|---|---|
| Muse Spark 1.3 FreeMeta Muse Spark via OpenCode Zen | $0 (promo window) | n/a — promo, not a plan | $0.00free | Limited-time promo; exact-string access route; unpublished throughput caps | None — promo ends on the provider's schedule | 2026-09-07 |
| NVIDIA K3 FreeNVIDIA Build | $0 (dev / prototyping) | n/a — free tier | $0.00free | Account-variable quota; dev/prototyping scope only | None — account limits govern | 2026-09-06 |
| Kimi K3 MembershipMoonshot AI | $19–$199/mo by tier | ≈$15–$159/mo prepaid | $32.50-59% vs PAYG ref | One shared credit pool; 5-hour and weekly usage controls | Monthly at list price; annual prepaid lowers effective cost | 2026-09-06 |
| Copilot ProGitHub | $10/user/mo (monthly billing; annual price for Pro not publ. … | see offer page | — | Included usage: $15/mo total credits on Pro (1,500 credits at $0 | Billed monthly per user | 2026-09-07 |
| ChatGPT PlusOpenAI | $19.99/mo in the official US App Store IAP list; the Plus h. … | see offer page | — | Message caps: per 5-hour windows and weekly allowances | Monthly auto-renew at the standard rate unless canceled | 2026-09-07 |
| Google AI ProGoogle | $19.99/mo (annual billing is offered but the annual price i. … | see offer page | — | Usage vs free tier: '4x access to Gemini' (footnoted 'in comparison to non-AI subscribers') | Charged at the beginning of each billing cycle | 2026-09-07 |
| Claude ProAnthropic | $20/mo billed monthly; annual is $200 billed up front ('$17. … | see offer page | — | Session: 'at least 5x' Free usage per rolling 5-hour window (exact token thresholds per model: . … | Auto-renews monthly (or yearly on annual, which 'is billed once up front for the year') | 2026-09-07 |
| Cursor ProCursor | $20/mo (annual USD price not published; a Monthly/Yearly to. … | see offer page | — | Hobby (free): 'Limited Agent requests', no card, Access to Composer — exact request counts not . … | Auto-renews monthly (or yearly where chosen — the yearly dollar figure is not published) from t. … | 2026-09-07 |
| Perplexity ProPerplexity | $20/mo; annual billing exists but its price is not publishe. … | see offer page | — | Pro Search: numeric Pro cap not published (Free = 3/day) | Monthly auto-renew at the standard rate unless canceled | 2026-09-07 |
| Per-100-task figures are illustrative (100k tokens/task reference; PAYG reference $0.80/task = $80 per 100 tasks, full-width bar). Basis per row shown under the figure. Snapshot 2026-09-06; re-verify at official terms — methodology | ||||||
Muse Spark 1.3 FreeMeta Muse Spark via OpenCode Zen
NVIDIA K3 FreeNVIDIA Build
Kimi K3 MembershipMoonshot AI
Copilot ProGitHub
ChatGPT PlusOpenAI
Google AI ProGoogle
Claude ProAnthropic
Cursor ProCursor
Perplexity ProPerplexity
Set your monthly task count and tokens per task to see whether pay-as-you-go or a flat subscription costs less at your volume. The flat-plan price is an input too — it defaults to the $39/mo Kimi Allegretto tier, so the crossover here matches the ≈49-task figure in the crossover story. The full method is documented in effective cost per task, explained.
For writing work, count your drafts per month and enter the number here: about 24 drafts clears the $19 Moderato month, about 49 clears the $39 Allegretto month — pick the cheapest tier your count clears. Tier details: Kimi K3 membership. More ways to run the numbers: the calculators page.
Inputs
Reference token blend (metered side)$8.00/M · $0.800/task
Premium ref is the illustrative reference blend the calculator defaults to — $0.800/task at your token size; a premium frontier API typically costs this much.
Lowest-cost option at these inputs
Flat subscription — $39/mo
$96.00/mo pay-as-you-go vs $39.00/mo flat · crossover 49 tasks/mo · $57.00/mo difference
The $39 default is the Kimi Allegretto tier — details in the tracked offer or the crossover story.
Verified Sep 6–7 2026 · the URL updates as you drag, so results are shareable.
Open the matching offer details with these inputsPay-as-you-go at $0.80/task$96.00/mo
$39 flat plan (Kimi Allegretto)$39.00/mo
Defaults: the $39/mo Allegretto tier vs $0.80/task pay-as-you-go crosses over at ≈49 tasks/mo (100k tokens/task). That default $0.80/task equals a premium-class $8 per 1M blended-token rate — pick a preset chip above if your traffic is K3- or flash-class, where metered cost per task is much lower and break-even moves accordingly. The per-task rate scales with the tokens slider at the selected $8.00 per 1M blend — your mix will differ. $40 is the illustrative reference basket vs pay-as-you-go; $39 is the actual Kimi Allegretto tier — break-even is ~50 tasks/mo on the basket, ~49 on Allegretto.
Break-even, in one chart
Metered billing climbs with every task (about $0.80 each at the reference rate); the $39 Allegretto month holds its price. The lines cross at ≈49 tasks/mo — annual billing moves the crossing to ≈39. The full four-step chart, ledger, and your-own-volume slider are on the crossover page.
27 routes with both a token price and a cited score · 6 on the frontier
Claude Fable 5.1
$20.00/M · 53 II · AA, accessed 2026-09-07
Claude Opus 5
$10.00/M · 51 II · AA, accessed 2026-09-07
GPT-5.6 Sol
$8.00/M · 47 II · AA, accessed 2026-09-07
GLM-5.3
$2.15/M · 44 II · AA, accessed 2026-09-07
GLM-5.3-Flash
$0.12/M · 42 II · AA, accessed 2026-09-07
Qwen3.8 Max
$3.00/M · 45* II · AA, accessed 2026-09-07
Staircase marks the cost-intelligence frontier: nothing to its lower right offers more intelligence per dollar. Hollow markers are $0 routes (quota-limited, unranked). Data: Artificial Analysis (artificialanalysis.ai), Intelligence Index v4.3, accessed 2026-09-07; per-point citations link the exact source. Cost arithmetic is ours.
43 paid per-token routes · sorted by blended $/M
A “task” here means one finished piece of work — a draft, a summary, a fix — estimated at 100k tokens unless you pick another size. Full definition: effective cost per task, explained.
Route type
| # | Route | Type | List in / out, $/M | Blended $/M | Batch $/M | Est. $/task @ Baseline 100k | AA II v4.3 | Src |
|---|---|---|---|---|---|---|---|---|
| 1 | Z.ai GLM-5.3-FlashBlend priced at promo $0.075/$0.25 — list-price blend $0.2375/M · Promo ends 2026-09-09 24:00 UTC+8API (per token) · 2 caveats | $0.07 / $0.25 | $0.119$0.012/taskcache: Cached input $0 | — | 42 | srcDIRECTread 2026-09-07 | ||
| 2 | Z.ai GLM-4.7-FlashX / Flash / 4.5-FlashZero-priced tiersAPI (per token) · 1 caveat | $0.07 / $0.40 | $0.153$0.015/taskcache: not published (free models | — | 15* | srcDIRECTread 2026-09-07 | ||
| 3 | DeepSeek deepseek-v4-flashPeak/off-peak split - see hoursAPI (per token) · 1 caveat | $0.22 / $0.66 | $0.330$0.033/taskcache: Cache-hit input $0 | — | 35 | srcDIRECTread 2026-09-07 | ||
| 4 | Z.ai GLM-4.6V / 4.5V / GLM-OCR / 4.6V-FlashLadder row; models not splitAPI (per token) · 1 caveat | $0.30 / $0.90 | $0.450$0.045/taskcache: Cached input e | — | — | srcDIRECTread 2026-09-07 | ||
| 5 | OpenAI gpt-5.6-lunaAPI (per token) · 0 caveats | $0.20 / $1.20 | $0.450$0.045/taskcache: 0 | $0.225~$0.022/task | 38 | srcDIRECTread 2026-09-07 | ||
| 6 | MiniMax MiniMax-M2.7API (per token) · 0 caveats | $0.30 / $1.20 | $0.525$0.052/taskcache: Cache-read $0 | — | 30 | srcEXCERPTread 2026-09-07 | ||
| 7 | Google Gemini 3.1 Flash-LiteAPI (per token) · 0 caveats | $0.25 / $1.50 | $0.563$0.056/taskcache: Cached input $0 | — | — | srcDIRECTread 2026-09-07 | ||
| 8 | OpenAI gpt-5-mini / gpt-5-nanoTwo models on one rowAPI (per token) · 1 caveat | $0.25 / $2.00 | $0.688$0.069/taskcache: 90% off cached input | $0.344~$0.034/task | 17 | srcDIRECTread 2026-09-07 | ||
| 9 | Mistral Mistral Large (API)Full table not renderedAPI (per token) · 1 caveat | $0.50 / $1.50 | $0.750$0.075/taskcache: Up to 90% off input for repeated prompts (cached input) | $0.375~$0.037/task | 10 | srcEXCERPTread 2026-09-07 | ||
| 10 | Google Gemini 2.5 Flash / 2.5 Flash-LiteTwo models on one rowAPI (per token) · 1 caveat | $0.30 / $2.50 | $0.850$0.085/taskcache: Cached input $0 | — | — | srcDIRECTread 2026-09-07 | ||
| 11 | DeepSeek deepseek-v4-proPeak/off-peak split - see hoursAPI (per token) · 1 caveat | $0.66 / $1.98 | $0.990$0.099/taskcache: Cache-hit input $0 | — | 36 | srcDIRECTread 2026-09-07 | ||
| 12 | Meta (Llama) via Together AI Llama 3.3 70B / Llama 3 8B Instruct LiteLlama 4 pricing unverifiedAPI (per token) · 1 caveat | $1.04 / $1.04 | $1.04$0.104/taskcache: Cached input prices shown per model | — | 9 | srcDIRECTread 2026-09-07 | ||
| 13 | MiniMax MiniMax-M2.7-highspeedPriority multipliers raise priceAPI (per token) · 1 caveat | $0.60 / $2.40 | $1.05$0.105/taskcache: Cache-read $0 | — | — | srcEXCERPTread 2026-09-07 | ||
| 14 | Google Gemini 3 Flash PreviewPreview tierAPI (per token) · 1 caveat | $0.50 / $3.00 | $1.13$0.113/taskcache: Cached input $0 | — | — | srcDIRECTread 2026-09-07 | ||
| 15 | Alibaba (Qwen) via Together AI Qwen3.6-Plus / Qwen3.5-397B-A17B / Qwen3.5-9B / Qwen3-235B-A22BResale pricing, not first-partyAPI (per token) · 1 caveat | $0.50 / $3.00 | $1.13$0.113/taskcache: Cached input $0 | — | — | srcDIRECTread 2026-09-07 | ||
| 16 | xAI grok-build-0.1 (coding)Whole-request repricing above 200kAPI (per token) · 1 caveat | $1.00 / $2.00 | $1.25$0.125/taskcache: Cached input $0 | — | — | srcDIRECTread 2026-09-07 | ||
| 17 | Google Gemini 3.8 / 3.7 / 3.6 FlashScheduled price increase 2027-01-01API (per token) · 1 caveat | $0.75 / $3.75 | $1.50$0.150/taskcache: Cached input $0 | ~$0.750~$0.075/task | 41 | srcDIRECTread 2026-09-07 | ||
| 18 | Z.ai GLM-5 / 4.7 / 4.6 / 4.5 ladderLadder row; models not splitAPI (per token) · 1 caveat | $1.00 / $3.20 | $1.55$0.155/taskcache: Cached input e | — | — | srcDIRECTread 2026-09-07 | ||
| 19 | xAI grok-4.3Whole-request repricing above 200kAPI (per token) · 1 caveat | $1.25 / $2.50 | $1.56$0.156/taskcache: Cached input $0 | — | — | srcDIRECTread 2026-09-07 | ||
| 20 | xAI grok-4.20-0309 (reasoning / non-reasoning / multi-agent)Logprobs unsupportedAPI (per token) · 1 caveat | $1.25 / $2.50 | $1.56$0.156/taskcache: Cached input $0 | — | — | srcDIRECTread 2026-09-07 | ||
| 21 | Moonshot AI Kimi K2.7 Code / K2.7-highspeed / K2.6 (API)Cache-hit rates differ - see notesAPI (per token) · 1 caveat | $0.95 / $4.00 | $1.71$0.171/taskcache: Cache-hit $0 | — | 29* | srcDIRECTread 2026-09-07 | ||
| 22 | Anthropic Haiku 4.5API (per token) · 0 caveats | $1.00 / $5.00 | $2.00$0.200/taskcache: 0 | — | 18 | srcDIRECTread 2026-09-07 | ||
| 23 | Z.ai GLM-5.3 / GLM-5.2 / GLM-5.1Context windows unpublishedAPI (per token) · 1 caveat | $1.40 / $4.40 | $2.15$0.215/taskcache: Cached input $0 | — | 44 | srcDIRECTread 2026-09-07 | ||
| 24 | DeepInfra / Hyperbolic / Novita / Lambda Third-party hosting (observed, not rowed)Unverified third-party figureCredits / prepaid · 1 caveat | $1.74 / $3.48 | $2.17$0.217/task | — | 36 | srcUNCERTAINread 2026-09-07 | ||
| 25 | xAI grok-4.6Whole-request repricing above 200kAPI (per token) · 1 caveat | $2.00 / $6.00 | $3.00$0.300/taskcache: Cached input $0 | — | 44 | srcDIRECTread 2026-09-07 | ||
| 26 | xAI grok-4.5Whole-request repricing above 200kAPI (per token) · 1 caveat | $2.00 / $6.00 | $3.00$0.300/taskcache: Cached input $0 | — | — | srcDIRECTread 2026-09-07 | ||
| 27 | Alibaba (Qwen) Qwen official API (Model Studio)Region/term tiers - see notesAPI (per token) · 1 caveat | $2.00 / $6.00 | $3.00$0.300/taskcache: 10% of input price, i | — | 45* | srcDIRECTread 2026-09-07 | ||
| 28 | Alibaba (Qwen) via Together AI Qwen3.8-2.4T-A95B / Qwen3.8 Flash / Qwen3.7-Max / Qwen3.7-PlusResale pricing, not first-partyAPI (per token) · 1 caveat | $2.00 / $6.00 | $3.00$0.300/taskcache: Cached input shown per model (e | — | 40 | srcDIRECTread 2026-09-07 | ||
| 29 | Google Gemini 3.5 Flash / 3.5 Flash-LiteTwo models on one rowAPI (per token) · 1 caveat | $1.50 / $9.00 | $3.38$0.338/taskcache: Cached input $0 | — | — | srcDIRECTread 2026-09-07 | ||
| 30 | Google Gemini 2.5 ProLong-context surcharge above 200kAPI (per token) · 1 caveat | $1.25 / $10.00 | $3.44$0.344/taskcache: Cached input $0 | $1.72~$0.172/task | 17 | srcDIRECTread 2026-09-07 | ||
| 31 | OpenAI o-series (o3 / o3-pro / o4-mini / o1 / o1-pro)Five models on one rowAPI (per token) · 1 caveat | $2.00 / $8.00 | $3.50$0.350/taskcache: Differs by model: o3 75% off | $1.75~$0.175/task | 20* | srcDIRECTread 2026-09-07 | ||
| 32 | OpenAI gpt-4.1 / 4.1-mini / 4.1-nanoThree models on one rowAPI (per token) · 1 caveat | $2.00 / $8.00 | $3.50$0.350/taskcache: 75% off cached input | $1.75~$0.175/task | — | srcDIRECTread 2026-09-07 | ||
| 33 | Anthropic Sonnet 5API (per token) · 0 caveats | $2.00 / $10.00 | $4.00$0.400/taskcache: 0 | — | 38 | srcDIRECTread 2026-09-07 | ||
| 34 | OpenAI gpt-5.6-terraAPI (per token) · 0 caveats | $2.00 / $12.00 | $4.50$0.450/taskcache: 0 | $2.25~$0.225/task | 42 | srcDIRECTread 2026-09-07 | ||
| 35 | Google Gemini 3.1 Pro PreviewLong-context surcharge above 200k · No free tierAPI (per token) · 2 caveats | $2.00 / $12.00 | $4.50$0.450/taskcache: Cached input $0 | — | — | srcDIRECTread 2026-09-07 | ||
| 36 | OpenAI gpt-5.2 / gpt-5.1Two models on one rowAPI (per token) · 1 caveat | $1.75 / $14.00 | $4.81$0.481/taskcache: 90% off cached input | $2.41~$0.241/task | — | srcDIRECTread 2026-09-07 | ||
| 37 | Moonshot AI kimi-k3 (API)Taxes excludedAPI (per token) · 1 caveat | $3.00 / $15.00 | $6.00$0.600/taskcache: Cache-hit input $0 | — | 44 | srcDIRECTread 2026-09-07 | ||
| 38 | OpenAI gpt-5.6-solPromo expiry 2026-11-21API (per token) · 1 caveat | $4.00 / $20.00 | $8.00$0.800/taskcache: 0 | $4.00~$0.400/task | 47 | srcDIRECTread 2026-09-07 | ||
| 39 | Anthropic Opus 5API (per token) · 0 caveats | $5.00 / $25.00 | $10.00$1.000/taskcache: 0 | — | 51 | srcDIRECTread 2026-09-07 | ||
| 40 | OpenAI gpt-5.5 / 5.4 / 5.4-mini / 5.4-nanoLadder row; models not splitAPI (per token) · 1 caveat | $5.00 / $30.00 | $11.25$1.125/taskcache: 90% off cached input (pro variants: no cached rate published) | $5.63~$0.563/task | 27 | srcDIRECTread 2026-09-07 | ||
| 41 | Anthropic Fable 5.1High output rateAPI (per token) · 1 caveat | $10.00 / $50.00 | $20.00$2.000/taskcache: 0 | $10.00~$1.000/task | 53 | srcDIRECTread 2026-09-07 | ||
| 42 | OpenAI gpt-6-astra (newest flagship)Newest flagship; premium rateAPI (per token) · 1 caveat | $10.00 / $50.00 | $20.00$2.000/taskcache: 0 | $10.00~$1.000/task | 53 | srcDIRECTread 2026-09-07 | ||
| 43 | OpenAI gpt-5-pro / 5.2-pro / 5.5-proThree models on one row; no cache rateAPI (per token) · 1 caveat | $15.00 / $120.00 | $41.25$4.125/taskcache: not published (no cached rate in table) | $20.63~$2.063/task | — | srcDIRECTread 2026-09-07 |
Blended $/M = (3 x input + 1 x output) / 4 at the route's current published price — list, or launch-promo price while a promo runs, with promo-priced rows captioning their list-price blend. Our arithmetic, not a provider figure. The $/task column estimates a 100,000-token task on that blended rate. Batch $/M = blended $/M x (1 − the provider's published batch discount) — −50% halves the blend — with batch $/task at the same task size; it appears only where the provider publishes a batch rate, and a dash means no published batch modifier, not zero. Cache, off-peak, and residency adjustments stay out of both figures and are noted per row. AA Intelligence Index v4.3 values are quoted, not measured, by us: every cell links its exact source row on artificialanalysis.ai (accessed 2026-09-07); * marks AA's own estimate flag. A dash means no public score. Our own weighted ranking is deliberately not shown — see methodology.
129 routes · 24 providers — 39 subscriptions · 48 api (per token) · 10 credits / prepaid · 18 coding tools · 14 free / promos
The raw ledger behind every ranking on this site — including rows we could not verify, labeled UNCERTAIN rather than priced from memory. Browse it filterable on the full route list, or go company by company via the provider pages.
Whether a subscription can be gifted, what refund windows mean for a prepay gift, and what to buy instead when it can't — for each tracked offer.
Read moreThe definition, the $40 / 120-task / $0.80 worked example, and the 50-task break-even, with the calculator.
Read moreAllegretto — the $39/month middle Kimi tier — at ≈$31/mo effective annual ($372 upfront): the prepay conditions and a checklist.
Read moreOne page per provider: every route we track, its price, and its evidence label — with cross-links to the leaderboard.
Read moreEffective-cost arithmetic, the quoted-score citation policy, staleness gates, and the data schema behind the tables.
Read moreCost data verified Sep 6–7 2026 · intelligence scores quoted from Artificial Analysis, accessed 2026-09-07 — verification log