Cheapest Fast LLM APIs (October 2026): Haiku 5.5, Luna, Qwen, DeepSeek, MiniMax

Updated 2026-10-26: We rebuilt this comparison around Claude Haiku 5.5, MiniMax M3.1 Flash, DeepSeek's V4.1-Flash-1017 snapshot, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Qwen 3.8 Flash and Muse Spark 1.3 Contributor stay in, with prices re-checked today. We removed Muse Spark 1.3 Standard, which at $1.25/$4.25 is outside the cheap tier, and GLM-5.3-FlashX, a paid speed tier that Artificial Analysis doesn't benchmark. We also added a note on Gemini 3.7 Flash and 3.5 Flash, which Google deprecated on 8 October and now reroutes. Scores and cost per task are from Artificial Analysis on 26 October 2026. Speeds are a same-day snapshot.

Three changes have reshaped the bottom of the LLM API market since our 5 October revision. Anthropic's pricing page lists Claude Haiku 5.5 at $0.10/$0.50, the same per-token price as GPT-6 Luna and a tenth of Haiku 4.5's $1/$5. DeepSeek moved its deepseek-flash endpoint to a new snapshot, DeepSeek-V4.1-Flash-1017, which now leads the cheap tier on Artificial Analysis's Intelligence Index. And MiniMax shipped M3.1 Flash at a flat $0.15/$0.60, with no peak-hour multiplier and no context cliff.

Separately, Google cleaned up its Flash lineup. gemini-3.7-flash and gemini-3.5-flash were deprecated on 8 October 2026. Their endpoints still answer, but requests are silently routed to Gemini 3.8 Flash and 3.6 Flash. If your cost model still has rows for those two models, you are being billed for different models than you think.

This page stays scoped to small, fast, high-volume models: classification, extraction, routing, cheap agent steps and bulk document work. If you are comparing frontier models on cost per task, that analysis lives in our cheapest near-frontier AI models breakdown. Below you will find the current price table, the cliffs and conditions attached to each headline rate, independent speed and benchmark data, three worked workload costs, a short note on the expensive-but-fast option, and a decision table.

What are the cheapest fast LLM APIs in October 2026?

All prices are USD per 1M tokens, standard (synchronous) tier, read from each vendor's pricing page on 26 October 2026. The blended column is a 3:1 input-to-output mix, which is roughly what retrieval-heavy production traffic looks like: (3 × input + output) ÷ 4.

* Alibaba's billing page marks qwen3.8-flash as eligible for a context-caching discount but does not print the cached rate in its price table, so we leave it out of the cache-heavy workload below.

Batch rates are 50% lower where offered. That gives Haiku 5.5 $0.05/$0.25 below 100K and $0.25/$1.25 above it, and Luna $0.05/$0.005/$0.25 on Batch/Flex below 272K. Gemini 3.8 and 3.6 Flash are $0.375/$1.875 on Batch and Flex until the January doubling, and Gemini 3.5 Flash-Lite is $0.15/$1.25. Google's output prices include thinking tokens.

The verified context windows are 1M tokens for Haiku 5.5 (128K max output), 1,050,000 for Luna (922K max input, 128K max output), 1M for deepseek-flash (384K max output), 1M for MiniMax M3.1 Flash, 1M for Qwen 3.8 Flash (Alibaba prices a single 0–1M input band) and 1,048,576 for Muse Spark 1.3.

If your cost model still has rows for Gemini 3.7 Flash or Gemini 3.5 Flash, the next section explains what their deprecation means for live traffic. GLM-5.3-Flash stays as a reference row: its rate of $0.15 / $0.03 cached / $0.50 is unchanged, and Z.ai still lists cached-input storage as "Limited-time Free" with no end date. On Artificial Analysis it scores 42 at $0.28 per task, so DeepSeek 1017 and MiniMax M3.1 Flash beat it on both axes. Our GLM-5.3-Flash guide covers it in depth.

What happened to Gemini 3.7 Flash and Gemini 3.5 Flash?

Google's changelog records both as deprecated on 8 October 2026. Neither endpoint is shutting down. According to Google's deprecations page, each "automatically routes API requests to" its replacement:

• gemini-3.7-flash (GA 13 August 2026) now routes to gemini-3.8-flash.
• gemini-3.5-flash (GA 19 May 2026) now routes to gemini-3.6-flash.

Neither model can be selected in Google AI Studio any more. In practice, a service that still sends gemini-3.7-flash keeps working, but it is now running a different model with different behaviour. Evals you ran against 3.7 Flash no longer describe production. Pin the replacement ID explicitly and re-run your evals.

Gemini 3.6 Flash deserves a flag of its own. It costs the same per token as 3.8 Flash ($0.75/$3.75, doubling in January) but scores lower on Artificial Analysis: 34 against 41. Google describes 3.6 Flash as "balancing speed and multimodal capabilities across general agentic and everyday tasks," and 3.8 Flash as "our most intelligent Flash model." If you were auto-migrated from 3.5 Flash to 3.6 Flash, moving to 3.8 Flash costs nothing extra per token. It does cost more per task: on AA's suite, 3.8 Flash came to $3.49 per task against 3.6 Flash's $1.60, because it generated 170M output tokens against 90M. Artificial Analysis's 3.6 Flash page currently labels the model deprecated, but Google's own deprecations page lists no shutdown date for gemini-3.6-flash. Treat Google's page as authoritative.

What strings are attached to each headline price?

Most rates in the table come with a condition. These are the ones that change the bill.

Claude Haiku 5.5: a 100K-token prompt cliff

Haiku 5.5's $0.10/$0.50 applies only to requests whose prompt is 100,000 tokens or less. Above that, every rate goes up 5×: $0.50 input, $0.05 cache reads, $2.50 output. Anthropic's pricing page is specific about how this works. Prompt length "counts all of its input tokens, including cache reads and cache writes," and "a request over the threshold pays the higher prices even when part of its prompt is a cache hit." Each request is priced on its own, so the cliff applies per call, not per conversation.

That is a much lower threshold than Luna's, and it matters for agents. A coding agent whose context grows past 100K tokens mid-session moves from one of the cheapest tiers in this table to one priced above DeepSeek's peak rate, and the change happens without any warning. All other Claude 4.6-and-later models bill their full 1M window at a flat rate. Haiku 5.5 is the named exception. The Claude Haiku 5.5 guide covers how to keep context trimmed.

GPT-6 Luna: a 272K surcharge on the full request

OpenAI's model page states: "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. It is a gentler cliff than Haiku's (2× and 1.5× against 5×), and it sits nearly three times higher. Luna also supports reasoning.effort: none, which means it can run as a plain non-reasoning model with no thinking-token overhead. OpenAI also sells a Fast tier for Luna (Priority processing was renamed Fast mode on 30 July 2026) at $0.20 / $0.02 / $1.00. See our GPT-6 Luna guide for where the model breaks down.

Qwen 3.8 Flash: the price depends on the region

Alibaba Cloud Model Studio publishes two rates for qwen3.8-flash. The Beijing, Hong Kong, Frankfurt, Tokyo and Virginia "Global" deployments charge $0.113 input and $0.382 output. The Singapore "International" deployment, and the Virginia "US" deployment scope, charge $0.15 and $0.47. Neither has context tiering: one band covers prompts up to 1M tokens. If you are on the higher rate, switching endpoint is a 25% input saving and a 19% output saving for the same model, as long as your data-residency rules allow it. The Singapore listing also includes 1M free tokens, valid for 90 days.

Against Haiku 5.5 and Luna, Qwen's lower-rate regions charge $0.013 more for input and $0.118 less for output. Qwen comes out cheaper once output exceeds about 11% of input tokens (0.013 ÷ 0.118 ≈ 0.11). For very input-heavy work, such as the classification workload below, Haiku and Luna still win. We found no Artificial Analysis entry for the API-served qwen3.8-flash, so it has no row in the benchmark table. Our Qwen 3.8 Flash guide covers the model.

Muse Spark 1.3 Contributor: paid for with your data

Meta's Contributor mode runs Muse Spark 1.3 at $0.10 input, $0.002 cached and $0.20 output, against $1.25 / $0.15 / $4.25 in Standard mode. Meta describes the trade as "heavily discounted token prices, in exchange for permission to use your content to train future Meta models", and your content includes prompts, code, files, context and model responses. In Standard mode, your content is not used for training. Three more limits make Contributor a prototyping tier:

• Rate limits are 20× lower on requests. muse-spark-1.3-contributor allows 150 requests per minute and 3M tokens per minute. muse-spark-1.3 allows 3,000 RPM and 4M TPM. Limits apply per team, not per API key.
• No max reasoning effort. Meta's reasoning docs say the "max" level "isn't available with muse-spark-1.3-contributor." Muse Spark 1.3's Artificial Analysis score of 48 is measured at max effort on the Standard tier ($3.69 per task), so it does not describe what Contributor can do.
• Audio is degraded on 1.3. Meta says audio understanding in Muse Spark 1.3 "is not fully supported," and points audio work at Muse Spark 1.2 or Muse Voice Transcribe.

Do not send client code, customer PII or regulated data through Contributor mode. The Muse Spark 1.3 guide covers both tiers.

DeepSeek: a new snapshot, peak pricing and a deep cache discount

The deepseek-flash endpoint now serves DeepSeek-V4.1-Flash-1017. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, and DeepSeek's pricing page says they are served by V4.1-Flash-1017 and billed at the Flash price. One correction to our earlier version: DeepSeek's page now lists vision as supported on deepseek-flash, so it is no longer text-only.

Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. At peak, all rates double to $0.30 / $0.006 / $1.20. That is 7 hours a day on 5 days, or 35 of the week's 168 hours, so about 79% of the week bills at off-peak rates. Plan against off-peak and schedule batch jobs away from the peak windows. The cache-hit price of $0.003 per 1M is 50× cheaper than a cache miss, the steepest discount among the production tiers here. Artificial Analysis reports the 1017 weights on Hugging Face under MIT. More in our DeepSeek V4.1 Flash guide.

MiniMax M3.1 Flash: flat and simple

MiniMax M3.1 Flash costs $0.15 input, $0.015 cache read and $0.60 output on MiniMax's pricing page, with no context tier listed and no peak multiplier. That makes it the simplest model in this comparison to forecast. MiniMax describes it as a "frontier multimodal coding model with 1M context window and tunable thinking depth." Artificial Analysis lists the release date as 19 October 2026 and the weights as proprietary. Our MiniMax M3.1 Flash guide has setup details.

Gemini 3.8 and 3.6 Flash: prices double on 1 January 2027

Google puts the increase in the price cell itself: "$0.75 through December 31, 2026" and "$1.50 starting January 1, 2027." Output goes from $3.75 to $7.50, and context-cache storage goes from $0.50 to $1.00 per 1M tokens per hour. Batch, Flex and Priority all double too. Priority goes from $1.35/$6.75 to $2.70/$13.50. Any annual budget that uses today's Gemini rates is wrong for the 10 months after 31 December. Gemini 3.5 Flash-Lite shows no date-based change on the pricing page, and its cache storage is already $1.00 per 1M tokens per hour. The Gemini 3.8 Flash guide covers the model itself.

A cost function that respects the cliffs

Cliffs reprice the whole request, so per-token averages give the wrong answer. The pure-Python sketch below applies each vendor's rule as documented. It calls no API and ignores cache-write premiums and batch discounts.

# USD per 1M tokens, standard tier, as of October 2026.
# Cache-write premiums and batch discounts are ignored.
RATES = {
    "claude-haiku-5-5": {
        "cliff": 100_000,
        "low": {"in": 0.10, "cached": 0.01, "out": 0.50},
        "high": {"in": 0.50, "cached": 0.05, "out": 2.50},
    },
    "gpt-6-luna": {
        "cliff": 272_000,
        "low": {"in": 0.10, "cached": 0.01, "out": 0.50},
        "high": {"in": 0.20, "cached": 0.02, "out": 0.75},
    },
    "deepseek-flash-offpeak": {"in": 0.15, "cached": 0.003, "out": 0.60},
    "MiniMax-M3.1-Flash": {"in": 0.15, "cached": 0.015, "out": 0.60},
    "gemini-3.8-flash": {"in": 0.75, "cached": 0.075, "out": 3.75},  # through Dec 31, 2026
}


def request_cost(model, uncached_in, cached_in, out):
    """Cost of one request in USD. Cliffs reprice the whole request."""
    r = RATES[model]
    if "cliff" in r:
        prompt_tokens = uncached_in + cached_in  # cached tokens count toward the cliff
        r = r["high"] if prompt_tokens > r["cliff"] else r["low"]
    return (uncached_in * r["in"] + cached_in * r["cached"] + out * r["out"]) / 1_000_000


for m in RATES:
    print(f"{m:24} 90K prompt: ${request_cost(m, 90_000, 0, 2_000):.4f}"
          f"   150K prompt: ${request_cost(m, 150_000, 0, 2_000):.4f}")

# claude-haiku-5-5         90K prompt: $0.0100   150K prompt: $0.0800
# gpt-6-luna               90K prompt: $0.0100   150K prompt: $0.0160
# deepseek-flash-offpeak   90K prompt: $0.0147   150K prompt: $0.0237
# MiniMax-M3.1-Flash       90K prompt: $0.0147   150K prompt: $0.0237
# gemini-3.8-flash         90K prompt: $0.0750   150K prompt: $0.1200

Moving from a 90K prompt to a 150K prompt makes Haiku 5.5 8× more expensive: $0.010 becomes $0.080, because 150,000 × $0.50 + 2,000 × $2.50 = $0.075 + $0.005. On the same jump, Luna's cost only scales with the extra tokens. At 150K, Haiku costs 5× what Luna does and more than three times what DeepSeek or MiniMax does.

How fast are they, and how do they score?

All speed and benchmark figures below come from Artificial Analysis (AA), an independent benchmarking firm. None are vendor-reported. Scores are on the AA Intelligence Index v4.3.2, which aggregates 10 evaluations, including Terminal-Bench 4.0, SciCode, Humanity's Last Exam and GDPval-AA v2.1. Output speed is the median tokens per second on the vendor's own API. Cost per task is AA's measured cost per Intelligence Index task. We read each model's AA page on 26 October 2026, and each row is the variant AA lists.

Scores are AA's rounded integers. Speed and latency are a snapshot from 26 October and moved by several percent between our reads that day. Five things stand out.

DeepSeek 1017 is the new capability leader of the cheap tier. It scores 45, two points above Haiku 5.5 and MiniMax M3.1 Flash, at $0.26 per task. At about 1.2s to first token, it is second only to MiniMax's roughly 1.05s in this table.

Luna is still cheapest per task by a wide margin. At $0.07 per task, it costs about a third of Haiku 5.5 at max effort ($0.21) and about 27% of DeepSeek ($0.26), for 5 to 7 fewer index points. It emitted 140M output tokens on the index, against Haiku's 440M.

Haiku 5.5 at max effort thinks a lot. Its time to first token of almost six minutes and its 440M output tokens are the cost of running it at max effort. AA's leaderboard also measures lower effort settings. At high effort, Haiku 5.5 scores 38 at about 185 tok/s, roughly 26s to first chunk and $0.08 per task. That matches Luna-at-max on score and roughly on cost, with a much shorter wait. At medium effort (Haiku's API default), it scores 34 at $0.05 per task. For latency-sensitive work, set effort explicitly rather than accepting the default.

Gemini 3.8 Flash is bought for something other than price-performance. At $3.49 per task, it costs about 13× DeepSeek 1017 ($3.49 ÷ $0.26 ≈ 13.4) and about 50× Luna ($3.49 ÷ $0.07 ≈ 49.9), while scoring below DeepSeek, Haiku 5.5, MiniMax and GLM. Gemini 3.6 Flash costs less per task ($1.60) but scores 34.

For raw throughput, Gemini 3.5 Flash-Lite is in a class of its own. It ran at 429.3 tok/s on our read, about 1.8× Haiku 5.5's 234.4 (429.3 ÷ 234.4 ≈ 1.83), but it scores 22. Use it for high-volume, low-judgement streaming, not for reasoning.

Luna's time to first token of about 94s is a max-effort artefact too. On AA's leaderboard, Luna at non-reasoning effort scores 18 with 0.93s to first chunk at $0.01 per task. That is the configuration to use for classification and routing, where the extra index points do not help.

What do these models cost on three real workloads?

Monthly costs at standard rates as of October 2026, with no caching except where stated, cache-write premiums ignored, and Gemini at 2026 rates. Gemini cache-storage fees are excluded from every Gemini row. Thinking tokens bill as output, so the output totals below assume minimal or no reasoning. Qwen figures use the lower-rate regions, with the Singapore rate in brackets.

1. High-volume classification

10M requests a month at 400 input and 30 output tokens each: 4,000M input and 300M output tokens.

• Muse Spark 1.3 Contributor: $460 (4,000 × $0.10 + 300 × $0.20), but this workload doesn't fit. At 150 requests per minute, a 30-day month allows 150 × 60 × 24 × 30 = 6,480,000 requests.
• Claude Haiku 5.5: $550. That is 4,000 × $0.10 = $400, plus 300 × $0.50 = $150. Prompts are far below the 100K cliff.
• GPT-6 Luna: $550. Same rates, same total. Use reasoning.effort: none.
• Qwen 3.8 Flash: $566.60 (4,000 × $0.113 + 300 × $0.382). On Singapore, $741.
• GLM-5.3-Flash: $750 (4,000 × $0.15 + 300 × $0.50)
• DeepSeek V4.1 Flash 1017, off-peak: $780 (4,000 × $0.15 + 300 × $0.60). At peak it would be $1,560.
• MiniMax M3.1 Flash: $780, at any hour
• Gemini 3.5 Flash-Lite: $1,950 (4,000 × $0.30 + 300 × $2.50), or $975 on Batch
• Gemini 3.8 Flash: $4,125 (4,000 × $0.75 + 300 × $3.75), or $2,062.50 on Batch. From January 2027, $8,250.

Haiku 5.5 and Luna tie exactly on price for short-prompt work. Choose between them on tooling, data residency and eval results, not cost. This is also the workload where Qwen's lower output rate doesn't pay off: output is only 7.5% of input here, below the 11% break-even. Input-heavy traffic punishes Gemini 3.8 Flash the most, because its input rate is 7.5× Haiku's and Luna's.

2. A coding agent

1,000 sessions a month, each reading 2M tokens of context across its turns at an 80% cache-hit rate and writing 120K output tokens. That is 1,600M cached input, 400M uncached input and 120M output.

• Muse Spark 1.3 Contributor: $67.20 (1,600 × $0.002 = $3.20, plus 400 × $0.10 = $40, plus 120 × $0.20 = $24). Your agent's code and context become Meta training data.
• Claude Haiku 5.5, every turn under 100K: $116. That is 1,600 × $0.01 = $16, plus 400 × $0.10 = $40, plus 120 × $0.50 = $60.
• GPT-6 Luna, turns under 272K: $116, by the same arithmetic.
• DeepSeek V4.1 Flash 1017, off-peak: $136.80 (1,600 × $0.003 = $4.80, plus 400 × $0.15 = $60, plus 120 × $0.60 = $72). At peak, $273.60.
• MiniMax M3.1 Flash: $156 (1,600 × $0.015 = $24, plus $60, plus $72)
• GLM-5.3-Flash: $168 (1,600 × $0.03 = $48, plus $60, plus 120 × $0.50 = $60)
• Qwen 3.8 Flash: up to $271.84 before its unpublished cache discount (2,000 × $0.113 + 120 × $0.382)
• Gemini 3.5 Flash-Lite: $468 (1,600 × $0.03 = $48, plus 400 × $0.30 = $120, plus 120 × $2.50 = $300)
• Claude Haiku 5.5, every turn over 100K: $580. That is 1,600 × $0.05 = $80, plus 400 × $0.50 = $200, plus 120 × $2.50 = $300.
• Gemini 3.8 Flash: $870 (1,600 × $0.075 = $120, plus 400 × $0.75 = $300, plus 120 × $3.75 = $450). From January 2027, $1,740.

Google also bills context-cache storage by the hour: $0.50 per 1M tokens per hour for Gemini 3.8 Flash until the January doubling, and $1.00 for Gemini 3.5 Flash-Lite. How much that adds depends on how long you keep caches alive.

The Haiku 5.5 rows are the story here. It is the same model doing the same work, and the bill grows 5× depending on whether per-turn context stays under 100K. Real agents fall somewhere in between, so log per-request prompt length before committing to Haiku for agent loops. Capability also matters more here than anywhere else on this page. A loop that is 90% reliable per step is only about 35% reliable over ten steps (0.9¹⁰ ≈ 0.35). DeepSeek 1017's 45 against Luna's 38 is worth the extra $20.80 a month on this workload if it reduces failed sessions.

3. Long-document processing

50,000 documents a month at 300K input and 8K output tokens each: 15,000M input and 400M output tokens.

• Muse Spark 1.3 Contributor: $1,580 (15,000 × $0.10 + 400 × $0.20). Meta lists no long-context premium, but every document becomes training data.
• Qwen 3.8 Flash: $1,847.80 (15,000 × $0.113 + 400 × $0.382). On Singapore, $2,438.
• GLM-5.3-Flash: $2,450 (15,000 × $0.15 + 400 × $0.50)
• DeepSeek V4.1 Flash 1017, off-peak: $2,490 (15,000 × $0.15 + 400 × $0.60)
• MiniMax M3.1 Flash: $2,490
• GPT-6 Luna, unchunked: $3,300. Each 300K prompt is over 272K, so the bill is 15,000 × $0.20 + 400 × $0.75.
• Gemini 3.5 Flash-Lite: $5,500 (15,000 × $0.30 + 400 × $2.50), or $2,750 on Batch
• Claude Haiku 5.5, unchunked: $8,500. Each prompt is over 100K, so the bill is 15,000 × $0.50 + 400 × $2.50. On Batch it is $4,250.
• Gemini 3.8 Flash: $12,750 (15,000 × $0.75 + 400 × $3.75), or $6,375 on Batch

Chunk the same documents so that every request stays under 100K tokens (four 75K chunks, for example), and both Haiku 5.5 and Luna fall to $1,700: 15,000 × $0.10 + 400 × $0.50. That is the cheapest result in this section among tiers that don't train on your data. This assumes chunking adds no overlap or extra output, which in practice it usually does. Even so, how you window the input decides a 5× swing for Haiku and a 1.9× swing for Luna. If you cannot chunk, for example because the task needs whole-document reasoning, the flat-priced rows are the safe default: Qwen 3.8 Flash on a lower-rate region, then DeepSeek off-peak and MiniMax.

What if you need maximum speed regardless of cost?

Can you self-host any of these models?

Only a few, and that matters if your blocker is data governance rather than price.

Watch the GLM naming. GLM-5.3-Flash is MIT, but the full GLM-5.3 is not: Hugging Face lists its licence as "other", so get legal sign-off before treating the two as interchangeable.

Self-hosting DeepSeek 1017 is the only route in this table that combines the cheap tier's top index score with prompts no third party sees. At $0.15/$0.60 off-peak, though, the API is hard to beat on cost. Our GLM-5.3-Flash self-hosting walkthrough found that a month of 4× B200 rental buys roughly 38 billion output tokens of GLM-5.3-Flash API usage instead. Breaking even therefore takes about 15,000 output tokens per second, sustained for the whole month. Treat that as the order of magnitude for any model in this price band.

Which cheap, fast model should you pick?

Our summary for late October 2026: Haiku 5.5 and Luna have the cheapest input tokens; Qwen 3.8 Flash has the cheapest blended rate if your region allows it; Luna is the cheapest finished task; DeepSeek 1017 is the best capability per dollar; MiniMax M3.1 Flash is the easiest bill to forecast. Muse Spark Contributor undercuts all of them only because you pay with your data. Gemini Flash is bought for Google integration or throughput, and should already be priced at its 2027 rate. Behind all of it, almost every headline rate on this page depends on a condition: a prompt-length cliff, a region, a peak window, a training opt-in, a published doubling date or a silent reroute. Re-read the vendor pricing pages each quarter and log per-request prompt lengths so a cliff never catches you by surprise.

FAQ

What is the cheapest fast LLM API right now?

As of October 2026, among APIs that don't train on your data, Claude Haiku 5.5 and GPT-6 Luna have the lowest input rate at $0.10 input and $0.50 output per 1M tokens, for prompts up to 100K and 272K tokens respectively. Qwen 3.8 Flash is cheaper on a 3:1 blend ($0.113/$0.382 outside Singapore, $0.180 blended against $0.200). Per completed task, Luna is cheapest on Artificial Analysis at $0.07 per Intelligence Index task at max effort.

Is Muse Spark 1.3 Contributor the cheapest LLM API?

On sticker price, yes: $0.10 input, $0.002 cached and $0.20 output per 1M tokens. The catch is that Meta uses your prompts, files and responses to train future models. It is also limited to 150 requests per minute and can't use max reasoning effort, so it suits prototyping on non-sensitive data, not production.

Is Claude Haiku 5.5 cheaper than GPT-6 Luna?

Per token they are identical below 100K tokens, including $0.01 cached input. Between 100K and 272K, Luna is 5× cheaper on input and output, because Haiku reprices to $0.50/$2.50 while Luna stays at $0.10/$0.50. On AA, Haiku 5.5 at max effort scores higher (43 vs 38) but costs $0.21 per task against Luna's $0.07. At high effort, Haiku scores 38 at $0.08.

How does Claude Haiku 5.5's 100K pricing work?

Any request whose prompt exceeds 100,000 tokens, counting cache reads and cache writes, pays $0.50 input, $0.05 cache reads and $2.50 output per 1M tokens for the whole request. Each request is priced independently, so earlier requests in a conversation keep the rate they were billed at.

What happened to Gemini 3.7 Flash?

Google deprecated gemini-3.7-flash on 8 October 2026. The endpoint still accepts requests but routes them to gemini-3.8-flash. gemini-3.5-flash was deprecated the same day and routes to gemini-3.6-flash. Neither is selectable in Google AI Studio any more.

What changed with DeepSeek-Flash in October?

The deepseek-flash endpoint now serves DeepSeek-V4.1-Flash-1017, which Artificial Analysis dates to 17 October 2026. Pricing is unchanged at $0.15/$0.60 off-peak and $0.30/$1.20 at peak, with $0.003 cache hits. DeepSeek's page now lists vision as supported. On AA it scores 45, the highest in this comparison.

Does Gemini 3.8 Flash pricing change in 2027?

Yes. Google's pricing page lists $0.75 input and $3.75 output per 1M tokens through 31 December 2026, then $1.50 and $7.50 from 1 January 2027. Gemini 3.6 Flash follows the same schedule, and Batch, Flex, Priority and cache storage all double too.

What is GPT-6.1 Sol Ultrafast?

It is OpenAI's fastest service tier, enabled with service_tier: "ultrafast". It costs $12 input and $60 output per 1M tokens for GPT-6.1 Sol up to 272K, which is 6× Standard. OpenAI publishes no tokens-per-second figure for it, and Artificial Analysis does not benchmark it separately.

Sources

• Anthropic: Claude API pricing
• Anthropic: Models overview
• OpenAI: API pricing
• OpenAI: GPT-6 Luna model page
• OpenAI: GPT-6.1 Sol model page
• OpenAI: Ultrafast mode guide
• DeepSeek: Models and pricing
• MiniMax: Pricing
• MiniMax: Models
• Alibaba Cloud Model Studio: Billing
• Meta Model API: Pricing and rate limits
• Meta Model API: Models
• Meta Model API: Reasoning
• Google: Gemini API pricing
• Google: Gemini API deprecations
• Google: Gemini API changelog
• Google: Gemini models
• Z.ai: Pricing
• Hugging Face: zai-org/GLM-5.3
• Artificial Analysis: Models leaderboard
• Artificial Analysis: Intelligence Index methodology
• Artificial Analysis model pages: Claude Haiku 5.5, GPT-6 Luna, DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, Gemini 3.8 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, GLM-5.3-Flash, Muse Spark 1.3

If your team is building high-volume pipelines or agents on these models and needs engineers who can manage cost cliffs, caching and evals, Codersera can help you hire vetted remote AI developers.