Cheapest Fast LLM APIs in 2026: Real Cost Per Task

If you built a cost model for cheap LLM APIs before mid-September 2026, it is wrong. Two things changed and neither was announced loudly. GLM-5.3-Flash's launch promotion expired at 24:00 on 9 September 2026 (UTC+8) and its effective price doubled — Z.ai deleted the discount notice from its pricing page rather than publishing anything, which is why most of the internet still quotes the promotional rate. And on 22 September, OpenAI shipped GPT-6 Luna at $0.10/$0.50, undercutting GLM on input for the first time.

Separately, DeepSeek retired the model this comparison used to recommend, and Artificial Analysis moved to a new version of its Intelligence Index, which reset every score anyone was quoting. This page was rebuilt against vendor pricing pages and AA model pages read on 5 October 2026. Where a figure differs from what an aggregator reports, the primary source wins and we say so.

What are the cheapest fast LLM APIs in October 2026?

Every figure below was read off the vendor's own pricing page on 5 October 2026. USD per 1M tokens. The blended column is a 3:1 input-to-output mix, roughly what a retrieval-heavy production workload looks like.

Sources: Z.ai pricing, OpenAI pricing, DeepSeek pricing, Gemini API pricing, Alibaba Model Studio billing, Meta Model API pricing.

A note on the blended column: it ranks sticker prices, not value. Four of these rows are conditional — one expiry has already fired, one has a context cliff, one has a peak-hours multiplier, one doubles in January. The sections below unpick each, and cost per completed task reorders the table substantially.

Why did GLM-5.3-Flash's price double?

Because a launch promotion ended and nobody said anything.

GLM-5.3-Flash shipped on 26 August 2026 with a flat 50% discount. Z.ai's pricing page showed list rates struck through next to the discounted ones, plus an explicit note: "GLM-5.3-Flash is available at a 50% discount (strikethrough prices are list prices). The promotion ends at 24:00 on September 9, 2026 (UTC+8, Singapore time)." That put Flash at $0.075 input / $0.015 cached / $0.25 output.

On 10 September the discount lapsed. Z.ai published no blog post, no changelog entry and no deprecation notice — it removed the strikethrough prices and the note, leaving the list rates of $0.15 / $0.03 / $0.50 standing as if nothing had happened. There is no promotional label on the GLM-5.3-Flash row today at all.

The practical consequence is a quiet 2× increase for anyone who budgeted during the promotional window, and a great deal of stale content. If you see GLM-5.3-Flash quoted at $0.075 — in a comparison table, a cost calculator, or an AI assistant's answer — that figure stopped being true on 10 September 2026.

Two things that are not price changes, and are easy to confuse with one:

• Z.ai is running a quota campaign to 7 October 2026 giving paid Coding Plan users zero-quota or doubled-quota GLM-5.3-Flash usage between 23:00 and 09:00 Singapore time. It affects plan credit consumption, not API list prices.
• A second Coding Plan offer runs 25 September to 7 October 2026, charging all-day usage at the off-peak credit rate. Also plan credits, also not the API.

Cached-input storage is still marked "Limited-time Free" on Z.ai's pricing page, with no end date published. Given what just happened to the launch discount, treat that as a rate that can disappear without warning too.

Is GPT-6 Luna actually cheaper than GLM-5.3-Flash?

Yes, on both axes that matter — and this is the first time a frontier lab has undercut the Chinese open-weight models at the bottom of the market. The capability gap is real but much smaller than the price gap.

GLM-5.3-Flash is the better model — 3.7 index points better — but costs 2.3× as much to complete the same evaluation suite, a steep premium for a 9% capability gain. OpenAI positions Luna as "our most efficient model for focused, high-volume tasks," below GPT-6.1 Sol and GPT-6 Astra, and that is honest: Luna is the volume model, not the judgement model.

Where Luna wins outright is mechanical work at scale. It supports reasoning.effort: none, which GPT-6 Astra does not — so you can run it as a plain non-reasoning model with no thinking-token overhead, the cheapest and lowest-latency configuration anywhere in this comparison. Pair that with the $0.01 cached-input rate and a stable system prompt and classification becomes close to free. Our GPT-6 Luna guide covers where the tier breaks down.

What GLM's premium buys is specific, not general: MIT weights you can host yourself, image/video/file input, and no 272K price cliff. If none of those three matter, Luna is the cheaper buy.

One correction, because aggregators are getting it wrong: there is no "GPT-6 Luna Pro" model. OpenAI's docs list exactly one snapshot, gpt-6-luna. What some catalogues show as a separate Pro model is the pro reasoning mode in the Responses API — identical per-token rates, higher spend, because it burns more tokens to reach an answer.

How do they compare on benchmarks?

Before the table, a warning that invalidates most published comparisons of these models: Artificial Analysis has moved to Intelligence Index v4.3.2, which aggregates 10 evaluations. Scores from the earlier v4.1.1 are not comparable. Figures in the 50s and 60s that circulated for these models in August and September were v4.1.1 numbers; the current index places this whole cohort in the high 30s and 40s. Z.ai's own marketing claim of "scoring 57" for GLM-5.3-Flash, for instance, was measured against v4.1.1 — it is not a contradiction of the 41.8 below, it is a different ruler.

Measured on v4.3.2, read from each model's Artificial Analysis page on 5 October 2026:

Four things there matter more than the ranking.

Gemini 3.8 Flash scores below GLM-5.3-Flash — 40.9 against 41.8. Inside the noise of a composite benchmark, but it means Gemini's 5× token premium buys no measurable capability over the cheapest open-weight competitor. Whatever Gemini is worth paying for, it is not intelligence.

The whole cohort is within 10 points, from 38.1 to 48.1, across a 20× spread in blended price. There is no cheap-and-stupid versus expensive-and-smart divide inside the fast tier any more — one capability band, wildly different price tags.

DeepSeek-Flash is the latency winner by a distance: 1.1 seconds to first token against 3.3s for GLM and 13.5s for Gemini. If perceived responsiveness is the product requirement, that beats the index.

Luna's 96-second time to first token is an artefact of max reasoning effort, not a property of the model — at reasoning.effort: none it is fast. Read it as a warning about effort settings generally.

Two models could not be scored, and we are not going to invent numbers. GLM-5.3-FlashX is not listed on Artificial Analysis at all — a serving tier over the same Flash weights, documented by Z.ai at 200 tokens/s for 2.5× the price, with no independent measurement. And the API-served qwen3.8-flash is unscored; the 39.8 above belongs to Qwen3.8-Flash-Next, a separate open-weights release. One is not a proxy for the other. (AA's leaderboard and per-model pages also print slightly different speed figures; the numbers above are from the per-model pages.)

What is the real cost per completed task?

Per-token price is the wrong unit. A model that reasons more, retries more, or rambles more can be cheaper per token and more expensive per finished job. Artificial Analysis publishes the actual dollar cost of running its full evaluation suite against each model — the closest thing to a like-for-like cost-per-task figure that exists:

Three findings fall out, and they are the real answer to "what is cheapest".

First, GPT-6 Luna wins on cost per task, not just sticker price. $122.05 against GLM's $280.28 — a gap that reflects Luna being less verbose (144M output tokens against 181M) on top of its lower input rate. When the cheapest-per-token model is also the cheapest-per-task model, the usual "cheap tokens are a trap" warning does not apply. It applies to the models below it.

Second, DeepSeek V4.1-Flash is verbose enough to erase its price advantage. It emitted 253M output tokens — the most here, 75% more than Luna — which is why a model priced within a cent of GLM on input costs 1.7× more per task while scoring lower.

Third, Gemini 3.8 Flash costs 13.3× what Luna does per task for 2.8 more points, and against GLM it is 5.8× the cost for 0.9 points fewer. There is no reading of this table in which Gemini is bought for price-performance; it is bought for throughput, audio, or Google-shaped procurement.

One caveat that keeps this honest: Muse Spark 1.3's $2,000.35 and 48.1 are measured at max reasoning effort, which is available on the standard tier only. The cheap Contributor tier cannot reach that score — the most important asterisk on this page, and the next section explains why.

What are the strings attached to the headline prices?

Every headline rate here comes with a condition. A price table that omits them is actively misleading, and in three cases the condition is new since this page was first written.

GLM-5.3-Flash: the promotion already expired. Covered above. $0.15/$0.50 is the live rate; $0.075/$0.25 is history as of 10 September 2026. The small upside is that GLM is now the only model here whose price has already absorbed its bad news.

GPT-6 Luna: a 272K price cliff. The $0.10/$0.50 is a short-context rate. OpenAI's model page states the rule plainly: "Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request." Above the line you pay $0.20/$0.02/$0.75. Note "for the full request" — not marginal pricing on the excess, it reprices the entire call. Against an advertised 1,050,000-token window that is a live trap, not an edge case. On the plus side, there is no documented introductory period or end date on Luna's rates, and OpenAI does flag promotional pricing with explicit dates when it applies — so unlike Gemini, Luna's price is not on a known clock.

DeepSeek: the model you were comparing no longer exists. The biggest change since September. DeepSeek's pricing page now reads: "the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but the corresponding models have been retired, their requests are served by the DeepSeek-V4.1-Flash model and billed at the Flash price." Send deepseek-flash instead. The old $0.22/$0.66 off-peak rate is gone; the replacement is $0.15 input / $0.60 output off-peak, doubling to $0.30/$1.20 at peak — peak being 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Because that is only 7 hours across 5 days, roughly 79% of the week bills off-peak, so off-peak is the rate to plan against rather than a discount to chase. DeepSeek's headline feature is now its cache: $0.003 per 1M on a hit, 50× cheaper than a miss, the most aggressive spread of any vendor here. Retiring the vision variant also leaves DeepSeek-Flash text-only.

Qwen 3.8 Flash: the price depends which region you call. Alibaba publishes two rates for the same model. The Singapore ("International") endpoint is $0.15/$0.47; the Beijing, Hong Kong, Frankfurt, US-Virginia and Tokyo endpoints are $0.113/$0.382 — about 25% cheaper on input and 19% on output for identical weights, with no context tiering and no end date. If you are on the Singapore rate, changing an endpoint is the cheapest optimisation on this page. New accounts also get 1M free tokens, valid 90 days from activation.

Gemini 3.8 Flash: the price doubles on 1 January 2027. Re-verified today, and still the condition almost nobody notices. Google writes the increase into the price cell rather than a footnote: input is "$0.75 through December 31, 2026. $1.50 starting January 1, 2027" and output "$3.75 through December 31, 2026. $7.50 starting January 1, 2027." Every tier doubles with it — batch and flex $0.375/$1.875 → $0.75/$3.75, priority $1.35/$6.75 → $2.70/$13.50. The output column is also labelled "including thinking tokens," and Gemini 3.8 Flash thinks a lot, so that is the rate your reasoning bills at.

Muse Spark Contributor: paid for with your data, and capped three ways. Contributor is 12.5× cheaper on input and 21× cheaper on output than standard Muse Spark 1.3, and Meta states the trade plainly: "Heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models." The standard tier explicitly does not — "your prompts and completions are not used to train Meta models." Training is also the one AI data decision you cannot walk back: retention windows expire and logs get deleted, but material absorbed into model weights has no extraction procedure.

Three further limits make Contributor a prototyping tier rather than a production one, and they are easy to miss:

• Max reasoning effort is standard-tier only. Meta's docs say Muse Spark 1.3 "supports all reasoning effort levels, including the 'max' level for extended reasoning (available on Standard tier only)." The 48.1 index score is measured at max. Contributor cannot reach it, and Meta publishes no score for what it can reach. Any comparison pairing "$0.10/$0.20" with "48 index" — including the earlier version of this page — is wrong.
• Rate limits are 30× lower. Contributor allows 100 requests per minute and 3M tokens per minute; standard allows 3,000 RPM and 4M TPM. Limits apply per team, not per API key.
• Audio is degraded on 1.3. Meta's docs: "Audio understanding in Muse Spark 1.3 is currently not fully supported, and response quality for requests including audio content may be degraded." Meta points audio work at Muse Spark 1.2 or Muse Voice Transcribe instead.

Practical reading: do not put client code, customer PII, or anything under HIPAA/GDPR-style obligations through the Contributor tier.

What do these models cost on three real workloads?

Monthly costs at each vendor's current rate, no caching except where stated.

1. High-volume classification. 10M requests/month, 400 input and 30 output tokens each (4B in, 300M out):

• Muse Spark Contributor: $460 — but 100 RPM caps you at ~4.3M requests/month, so this workload does not fit
• GPT-6 Luna: $550 — Luna's best case, at reasoning.effort: none
• Qwen 3.8 Flash, Beijing endpoint: $567 — $741 on Singapore
• GLM-5.3-Flash: $750 — was $375 before the promotion expired
• DeepSeek-Flash, off-peak: $780 — $1,560 at peak
• Gemini 3.8 Flash, batch: $2,063 — $4,125 synchronous, and both double in January

Input-dominated work punishes Gemini hardest: its input premium is 7.5× Luna's. This is also the workload where Luna's lower index score is irrelevant — classification against a fixed enum does not need the extra 3.7 points. And note the Contributor rate limit disqualifies the cheapest row outright, which is the kind of thing a price table never shows you.

2. A coding agent. 1,000 sessions/month, each reading 2M tokens of context at 80% cache hits and writing 120K tokens:

• Muse Spark Contributor: $67
• GPT-6 Luna: $116
• Qwen 3.8 Flash, Beijing: $117 — $142 on Singapore
• DeepSeek-Flash, off-peak: $137 — its $0.003 cache-hit rate does the work here
• GLM-5.3-Flash: $168
• Gemini 3.8 Flash: $870, plus cache storage at $0.50 per 1M tokens per hour

Agent workloads are cache-read dominated, where these models diverge most violently — a 75× spread between Muse Contributor's $0.002 and Muse standard's $0.15. The caveat that overrides the table: agents are where small per-step quality gaps compound. A loop that is 90% reliable per step is 35% reliable over ten steps, and cheap steps that fail are not cheap. The 3.7 index points separating Luna from GLM-5.3-Flash are worth more here than anywhere else on this page.

3. Long-document processing. 50,000 documents/month at 300K input and 8K output each (15B in, 400M out):

• Muse Spark Contributor: $1,580
• Qwen 3.8 Flash, Beijing: $1,848 — $2,438 on Singapore
• GLM-5.3-Flash: $2,450
• DeepSeek-Flash, off-peak: $2,490
• GPT-6 Luna: $3,300 — not $1,700
• Gemini 3.8 Flash, batch: $6,375 — $12,750 synchronous

That Luna line is the 272K cliff in action, and it is the most useful number on this page. At 300K input tokens per document you are over the threshold, so the entire request reprices to $0.20/$0.75 and the cheapest-input model in the table becomes the second most expensive option. Chunk the same documents to stay under 272K and Luna drops to $1,700 — the cheapest no-strings result of all three workloads. Same model, same month, same volume, 1.9× difference, decided purely by how you window your inputs.

Can you self-host any of these models?

Some of them, and the licence details are messier than most coverage admits.

Two traps there. GLM-5.3-Flash is MIT but GLM-5.3 is not, and a lot of reporting has conflated them. The full model's LICENSE adds a clause: a licensee operating a Model-as-a-Service business whose aggregate revenue with its affiliates exceeds "10 billion US dollars … in total over any consecutive 12 months" must pass a Z.AI security review before commercial use. Two precision points, since this gets garbled in summary: the $10B test is on total revenue, not MaaS revenue, and it bites only if you also operate a MaaS business; and the consequence is a mandatory review whose scope Z.AI determines, not a fee or a ban. For anyone not running a ten-billion-dollar cloud it is inert — but the licence string is other, not mit, and if your legal team asks, the distinction matters. The second trap is naming: the API-served qwen3.8-flash and the open-weights Qwen3.8-Flash-Next are different artifacts, not one model in two delivery modes.

Self-hosting is the exit if your blocker is data governance rather than price — the only option where no third party sees your prompts. The arithmetic is unforgiving, though: our GLM-5.3-Flash self-hosting walkthrough works out that a month of 4× B200 rental buys roughly 38 billion output tokens of API usage instead, so breaking even needs about 15,000 output tokens per second sustained for a month.

Which cheap model should you actually pick?

The decision rule, in order:

1. Is the data sensitive? Rule out Muse Spark Contributor immediately. Consider self-hosting GLM-5.3-Flash or DeepSeek-Flash.
2. Need audio input? Gemini 3.8 Flash. Muse Spark 1.3 nominally accepts audio but Meta documents it as degraded; everything else is text-and-image or text-only.
3. Is the task mechanical — classification, extraction, routing, labelling? GPT-6 Luna at reasoning.effort: none, prompts under 272K. Cheapest per token and cheapest per task.
4. Does the task need judgement — agents, coding, multi-step work? GLM-5.3-Flash. 3.7 index points over Luna, no context cliff, and MIT weights as an escape hatch if the price moves again.
5. Is latency the product? DeepSeek-Flash — 1.1s to first token, and a $0.003 cache-hit rate for repeated context.
6. Calling Qwen from Singapore? Move to Beijing, Frankfurt, Virginia or Tokyo and save ~22% for nothing.
7. Planning past December? Reprice Gemini at $1.50/$7.50 before committing.

There is no free lunch in this table, only differently-shaped bills. The honest summary for October 2026: GPT-6 Luna is the cheapest capable fast LLM API on both sticker price and cost per completed task, provided prompts stay under 272K and 38 index points is enough. GLM-5.3-Flash is what you pay 2.3× more for when the work needs judgement, and it remains the best value in its capability band even after its price doubled. Gemini 3.8 Flash scores below GLM at 5.8× the cost per task — buy it for throughput or audio, not price-performance, and price it at next year's rate already. Muse Spark Contributor is a prototyping tier wearing a production tier's price tag.

The meta-lesson outlasts every number above: most of these prices move on a schedule or a condition — an expiry that already fired silently, a context cliff that reprices whole requests, a peak-hours multiplier, a doubling in January. Re-read your vendor's pricing page quarterly. Nobody will email you.

Deeper single-model breakdowns: GPT-6 Luna, GLM-5.3-Flash, Qwen 3.8 Flash, Gemini 3.8 Flash, and Muse Spark 1.3.

FAQ

What is the cheapest LLM API in 2026?

Among capable fast-tier models, GPT-6 Luna at $0.10 input and $0.50 output per 1M tokens, verified 5 October 2026. It is also cheapest per completed task — $122.05 to run Artificial Analysis's evaluation suite against GLM-5.3-Flash's $280.28. Muse Spark 1.3 Contributor is cheaper on a blended basis at $0.10/$0.20, but requires letting Meta train on your prompts and is capped at 100 requests per minute.

Did GLM-5.3-Flash's price go up?

Yes. Its 50% launch promotion expired at 24:00 on 9 September 2026 (UTC+8), taking it from $0.075/$0.25 back to its list rate of $0.15/$0.50 per 1M tokens. Z.ai published no announcement — it simply removed the discount notice from its pricing page — which is why many comparisons and cost calculators still quote the promotional figure.

Is GPT-6 Luna cheaper than GLM-5.3-Flash?

Yes, on both axes: $0.10 against $0.15 on input with identical $0.50 output rates, and $122.05 against $280.28 to complete Artificial Analysis's evaluation suite. GLM-5.3-Flash scores higher — 41.8 against 38.1 on Intelligence Index v4.3.2 — and has no context price cliff, where Luna reprices to $0.20/$0.75 for any prompt above 272K input tokens. Use Luna for classification and extraction, GLM for agents and coding.

What happened to DeepSeek V4-Flash?

It was retired. DeepSeek's pricing page states the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but the models behind them are gone, with requests now served by DeepSeek-V4.1-Flash under the model name deepseek-flash. The old $0.22/$0.66 off-peak rate no longer exists; the current rate is $0.15/$0.60 off-peak and $0.30/$1.20 at peak. Retiring the vision variant also leaves DeepSeek-Flash text-only.

Does Gemini 3.8 Flash pricing change in 2027?

Yes, it doubles. Google writes the increase into the price cell itself: input goes from $0.75 to $1.50 and output from $3.75 to $7.50 per 1M tokens on 1 January 2027. Every tier doubles with it — batch from $0.375/$1.875 to $0.75/$3.75, priority from $1.35/$6.75 to $2.70/$13.50. As of 5 October 2026 there is no indication the introductory period has been extended.

Why does Qwen 3.8 Flash have two prices?

Because Alibaba prices Model Studio by region. The Singapore "International" endpoint charges $0.15 input and $0.47 output per 1M tokens; the Beijing, Hong Kong, Frankfurt, US-Virginia and Tokyo endpoints charge $0.113 and $0.382 for the same model. There is no context-length tiering on either, so switching endpoints is a ~22% saving for no change in capability.

Is GLM-5.3 MIT licensed?

GLM-5.3-Flash and GLM-5.2 are genuinely MIT. The full GLM-5.3 is not — it ships under a bespoke "GLM-5.3 License" adding one condition: an organisation that operates a Model-as-a-Service business and whose aggregate revenue with affiliates exceeds $10 billion over any consecutive 12 months must pass a Z.AI security review before commercial use. Below that threshold the grant is MIT-like, but the licence string is other, not mit.

Why did all the Intelligence Index scores drop?

Artificial Analysis moved to Intelligence Index v4.3.2, which aggregates 10 evaluations. Scores are not comparable across index versions. Figures in the 50s and 60s circulating for these models are v4.1.1 numbers; on v4.3.2 the whole fast tier sits between 38 and 48. When a vendor cites a score, check which index version it was measured against before comparing it to anything.

Is DeepSeek-Flash pricing always the same?

No. DeepSeek charges peak rates of $0.30 input and $1.20 output per 1M tokens during 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays, and half that off-peak. Because peak covers only 7 hours across 5 days, roughly 79% of the week bills at the off-peak $0.15/$0.60 — plan against off-peak, not peak. Cache hits cost $0.003 per 1M, 50× cheaper than a cache miss.