Budget AI Coding Models, October 2026: Haiku 5.5 vs GPT-6 Luna

Last updated: October 8, 2026 - added Claude Haiku 5.5 (released October 7) in place of Haiku 4.5, a Haiku 5.5 vs GPT-6 Luna cost section, Anthropic's head-to-head benchmark table, and re-checked OpenAI, Anthropic, DeepSeek and Gemini pricing pages.

The cheapest AI coding models in October 2026 tie on list price: Claude Haiku 5.5 and GPT-6 Luna both list at $0.10 per million input tokens and $0.50 per million output, with $0.01 cache reads. The price tie breaks on two things. Haiku 5.5 only holds that price for prompts up to 100,000 tokens (it jumps to $0.50/$2.50 above), while Luna holds it to 272K and then rises to just $0.20/$0.75. And Claude's newer tokenizer counts about 30% more tokens than Haiku 4.5 did for the same text. On capability they do not tie: Anthropic's launch table, which includes a GPT-6 Luna column, puts Haiku 5.5 at 39.2% on Terminal-Bench 4.0 against Luna's 16.4%. Those are vendor-reported numbers from Luna's competitor, and Haiku 5.5 still sits well below Sonnet 5.5 (70.6%) on the same benchmark.

So for short, bounded calls under 100K tokens - classification, summaries, compaction, subagent steps - the two cost about the same, Anthropic's numbers lean Haiku, and your own golden set and your stack should settle it. For long agent context, Luna is several times cheaper. DeepSeek V4.1-Flash ($0.15/$0.60 off-peak) is still the open-weights floor, and Gemini 3.5 Flash ($1.50/$9.00) and the now-legacy Claude Haiku 4.5 ($1/$5) sit a full price tier above.

The Quick Comparison#

Prices verified October 8, 2026 on the OpenAI pricing page, the GPT-6 Luna and GPT-5.6 Luna model pages, the Claude Haiku 5.5 overview, DeepSeek pricing and Gemini pricing.

Legacy note: Claude Haiku 4.5 ($1 input, $5 output, $0.10 cache read) is still available, but Anthropic's launch post prices Haiku 5.5 90% lower for requests up to 100K tokens. New budget work on the Anthropic side should start on Haiku 5.5; the Haiku 5.5 migration guide covers the ten request changes.

The price spread across the tier is now wider than the capability spread on many tasks. Three models - GPT-6 Luna, Haiku 5.5 under 100K, and V4.1-Flash - sit at or near ten cents per million input tokens, 10 to 15x below Gemini 3.5 Flash. The interesting question is no longer "which is cheapest" but "how much agentic capability does each model actually bring to a loop, and where does each price break."

Haiku 5.5 vs GPT-6 Luna: Same Price, Different Bill#

Same sticker, different bill. Both models charge $0.10 input, $0.50 output and $0.01 cache reads for short prompts. They differ in where the price breaks and in how tokens are counted:

• Haiku 5.5 charges $0.50 input and $2.50 output for prompts over 100,000 tokens, with cache reads at $0.05 (model overview). That is 5x the short-prompt rate.
• GPT-6 Luna prices prompts with more than 272K input tokens at 2x input and cache rates and 1.5x output for the full request: $0.20 / $0.75 (GPT-6 Luna model page).
• Tokenizer: Anthropic's migration guide says the same input text produces approximately 30% more tokens on Haiku 5.5 than on Haiku 4.5. That 30% is measured against Haiku 4.5, not against OpenAI's tokenizer, so nobody has documented how Haiku 5.5 and Luna counts compare for the same text. Our Haiku 5.5 release guide covers the cliff and the tokenizer in depth.

Here is what that does to a bill. Our arithmetic at list prices, no caching, using three task shapes. The "same count" column is the like-for-like comparison. The "+30% sensitivity" column is a what-if, not a measurement: it inflates only the Haiku side by the documented Haiku 4.5 gap, which assumes Luna counts text the way Haiku 4.5 did. Count your own prompts on both APIs before trusting either column.

Worked example, Luna side: the 200K-in, 20K-out agent task stays under Luna's 272K line, so it is 0.2M x $0.10 + 0.02M x $0.50 = $0.02 + $0.01 = $0.030.

Worked example, Haiku side: the same task crosses Haiku's 100K line, so the whole request bills at the upper tier: 0.2M x $0.50 + 0.02M x $2.50 = $0.10 + $0.05 = $0.15. That is 5x Luna's cost at the same token count, purely from the price cliff. Under 100K the picture flips back to a tie: the 50K subagent step costs the same at equal counts.

The practical rule: if you size prompts with Haiku 4.5 counts, Haiku 5.5's cliff sits at about 77K of those tokens (100K divided by 1.3). Keep Haiku calls short and cached and it matches Luna on price; let context grow past the line and Luna wins by a multiple. Anthropic says about 90% of Haiku 4.5 requests already fell under 100K tokens, so many existing Claude workloads land on the cheap side. Long-running agent loops that replay whole histories usually do not, which is why prompt caching and context trimming matter more than the sticker here.

One Anthropic-specific lever: Max and Team plans now include a monthly API credit ($100 on Max 5x, $200 on Max 20x, up to $500 pooled on Team), usable on the Claude API but not on interactive Claude Code (details). On the 50K subagent shape above, $200 covers roughly 20,500 Haiku 5.5 calls with the 30% inflation applied, against about 2,700 of the same calls on Haiku 4.5 at $0.075 each. If you already pay for Max, that credit stretches furthest on Haiku 5.5.

Claude Haiku 5.5: Anthropic Moves to the Floor#

Until October 7, Anthropic did not compete on price in this tier. Claude Haiku 5.5 changes that: claude-haiku-5-5, a 1M-token context window, 128K max output, adaptive thinking with a default effort of medium, and Batch API at 50% off (model overview). Anthropic positions it for high-volume, latency-sensitive work - classification, routing, extraction and subagent tasks.

Its role in this tier is the same-provider pick with a much lower floor. If the rest of your stack is Claude, keeping the budget model with the same provider removes the wire-format and eval-surface differences that cross-provider routing introduces - and that choice no longer costs 10x on short prompts. The catch is the migration: fixed thinking budgets, non-default sampling parameters and assistant prefill all return 400 errors on Haiku 5.5, so it is not a drop-in model id swap. Our Haiku 5.5 release guide walks through each change.

DeepSeek V4.1-Flash: The Open-Weights Floor#

DeepSeek's September 10 release news describes the generation precisely: DeepSeek-V4.1-Flash is the smallest model in a new architecture family - a 552B-parameter MoE with asymmetric causal encoder-decoder design (8B active for input, 16B for output), native visual understanding, and a quarter of the previous KV cache footprint. It has a 1M context and 384K max output and is served as deepseek-flash; the legacy deepseek-v4-flash names are still accepted but are served by V4.1-Flash and billed at the Flash price.

As of October 8, 2026 the pricing page lists V4.1-Flash at $0.15 input, $0.60 output and $0.003 cache hits off-peak, doubling to $0.30 / $1.20 / $0.006 during peak hours (01:00-04:00 and 06:00-10:00 UTC, Monday through Friday, excluding Chinese public holidays; weekends are off-peak). The same page lists deepseek-v4-pro as its own model (DeepSeek-V4-Pro-0813) at $0.66 / $1.98 off-peak; this comparison tracks only the Flash tier.

The access story is strong: deepseek-flash supports tool calls, JSON output, the Responses API and an Anthropic-format endpoint, and DeepSeek publishes a one-line Codex setup script. The July 31 update analysis walks through what changed when the Flash line started handling whole loops, and the September re-verify covers the rename.

GPT-6 Luna: The Floor With the Longest Runway#

OpenAI's GPT-6 family landed September 22, and Luna is its high-volume workhorse: $0.10 per million input tokens and $0.50 output, cache reads at $0.01 and cache writes at $0.125 (verified October 8, 2026 on the GPT-6 Luna model page). It undercuts DeepSeek V4.1-Flash ($0.15/$0.60 off-peak) and GPT-5.6 Luna ($0.20/$1.20), and Haiku 5.5 now matches it only below 100K tokens.

Luna is a full agentic citizen: a 1,050,000-token context window, 128K max output, long-context rates of $0.20/$0.75 above 272K input tokens, and batch at 50% off ($0.05/$0.25). That long-context runway is its structural advantage in this tier. The open question is capability: OpenAI's model page lists no benchmark scores, and the only head-to-head numbers we found come from Anthropic's Haiku 5.5 launch (see the benchmarks section below), which favor Haiku. Evaluate it on your own golden set before you commit.

GPT-5.6 Luna: Now the Middle of the OpenAI Family#

OpenAI's July 30 announcement cut GPT-5.6 Luna from $1/$6 to $0.20/$1.20 - an 80% drop - with cache hits at $0.02 and long-context rates of $0.40/$1.80. Our breakdown of the announcement covered the reaction: a typical agent conversation of 50K input and 10K output tokens dropped to about $0.022.

GPT-6 Luna now does the same job at half the price, so GPT-5.6 Luna is a holdover rather than an entry point. It is still the recommended replacement for gpt-5-nano on OpenAI's deprecations page (gpt-5-mini maps to GPT-5.6 Terra), and it is the entry tier of the GPT-5.6 family. New OpenAI integrations should start on GPT-6.

Gemini 3.5 Flash: The Hosted Middle#

Gemini 3.5 Flash lists at $1.50 input and $9.00 output, with cache hits at $0.15 plus $1.00 per million tokens per hour of cache storage (verified October 8, 2026 on the Gemini pricing page). That is 15x Luna and Haiku 5.5 on input and 18x on output, and the storage fee means a cache you hold but rarely hit can cost more than no cache. Google's newer Gemini 3.8 Flash is listed on the same page at $0.75 / $3.75 through December 31, 2026, rising to $1.50 / $7.50 from January 1, 2027 - still well above the floor. Gemini 3.1 Pro Preview at $2/$12 is the step up if the tier runs out of ceiling, and our comparison of the Anthropic and Google options covers that boundary.

The Cost-Per-Task Math#

The frame that matters in this tier is dollars per completed task, not dollars per million tokens. A representative agentic coding task - 200K input tokens, 20K output, no cache - runs, at October 8, 2026 list rates (our arithmetic):

• GPT-6 Luna: $0.020 + $0.010 = $0.030
• DeepSeek V4.1-Flash (off-peak): $0.030 + $0.012 = $0.042
• GPT-5.6 Luna: $0.040 + $0.024 = $0.064
• Claude Haiku 5.5 (over-100K tier): $0.10 + $0.05 = $0.15 at the same token count ($0.195 in the +30% sensitivity case)
• Gemini 3.5 Flash: $0.30 + $0.18 = $0.48

The spread is about 16x across the tier for the same token count. Haiku 5.5 lands in the middle on this shape only because 200K crosses its cliff; trim the same task under 100K and it ties Luna. The catch is that token counts are not the same across models: models that fail more retry more, and retries compound input tokens fast. A model that completes the task in one attempt at $0.48 is cheaper than one that needs five attempts at $0.042. This is the argument for measuring completion rates before switching, and it is the reason cost-per-task analysis keeps beating sticker-price comparisons.

DeepSeek adds a scheduling wrinkle no other vendor in this tier has: off-peak rates are half of peak, so batching agent work into the off-peak window (including all weekend hours) is how its already-low numbers stay low.

Benchmarks: What Is Verified, What Is Not#

The one head-to-head in this tier comes from Anthropic. Its Haiku 5.5 announcement includes a GPT-6 Luna column:

A dash means Anthropic's table has no number for that cell. Read this as a competitor's measurement: Anthropic points to the Haiku 5.5 system card for how it ran its evaluations, the post does not describe the settings it used for the Luna runs, and OpenAI's Luna model page publishes no scores to check against. Taken at face value, Haiku 5.5 is more than twice Luna on Terminal-Bench 4.0 and well ahead on computer use, while the FrontierCode gap is narrow. Anthropic also says Sonnet 5.5 (70.6% on Terminal-Bench 4.0) and Opus 5.5 remain the better choices for complex agentic coding - a useful signal that Haiku belongs on bounded steps, not whole coding loops.

The rest of the tier does not line up. DeepSeek published agent numbers for its retired V4-Flash 0731 (82.7 on Terminal Bench 2.1, 54.4 on DeepSWE, vendor-reported with config documented), but on different benchmarks, so they cannot be compared with Anthropic's table. Its V4.1 benchmark tables are images in the release news, not text we can quote precisely. None of these numbers are independent.

What the tier does have in common: all five models are cheap enough that running your own golden-set evaluation is trivial. Our eval-before-switch day plan applies verbatim - build a golden set, run it against all five, compare completion rates and cost per task. At this tier the eval is the cheapest part of the migration.

Decision Guide by Workload#

• High-volume extraction and classification: GPT-6 Luna, Claude Haiku 5.5 or DeepSeek V4.1-Flash. All three are under a cent per short call; pick by stack - OpenAI's Responses API and tool-calling (Luna), Anthropic's API and the Max plan credit (Haiku 5.5), or MIT open weights you can self-host (DeepSeek).
• Subagents and short tool steps under 100K tokens: Claude Haiku 5.5 if you are on Anthropic, GPT-6 Luna if you are on OpenAI. At these sizes the two tie on price, so capability and wire compatibility decide. If you are stack-neutral, Anthropic's vendor-reported table favors Haiku 5.5 on agentic coding and computer use; confirm that on your own golden set before routing traffic.
• Long-context agent loops (over 100K tokens per request): GPT-6 Luna. Its price only steps up past 272K, and then to $0.20/$0.75, while Haiku 5.5 bills 5x above 100K.
• Self-hosted or data-sensitive: DeepSeek V4.1-Flash. MIT open weights, 1M context, and the only model in this tier you can run on your own hardware.
• Everything else: measure. The 16x spread between cheapest and most expensive means the wrong pick by capability costs you in retries, not sticker price.

When to Skip the Budget Tier#

The budget tier fails in two ways worth naming. First, long-horizon autonomy: if your tasks run hours and tolerate no mid-task regression, the frontier tier exists for that - and the frontier is cheap too now, with GPT-6.1 Sol at $2/$10, so the "step up" is no longer a 10x price jump. Second, output-heavy reasoning: Luna's and Haiku 5.5's $0.50 output and Flash's $0.60 are still far below the frontier's $10-$50, but long reasoning traces on a budget model are where the quality gap shows up most.

The other skip signal is organizational: if your team already standardizes on one provider for compliance, support, or eval reasons, the cross-provider savings of this tier may not clear the bar - though with Haiku 5.5 matching Luna's list price under 100K, Anthropic shops now pay far less for staying put. The model routing strategies post covers when routing between tiers pays and when it does not.

Changelog#

• October 8, 2026: Claude Haiku 5.5 (October 7) replaces Haiku 4.5 in the comparison; new Haiku 5.5 vs GPT-6 Luna section and Anthropic's head-to-head benchmark table; GPT-5.6 Luna row completed (1.05M context, 128K output); DeepSeek, OpenAI and Gemini rates re-checked.
• September 26, 2026: GPT-6 Luna entered at $0.10/$0.50; DeepSeek's September 10 repricing cut the Flash tier to $0.15/$0.60 off-peak and renamed it deepseek-flash.
• August 15, 2026: DeepSeek's peak/off-peak policy took effect August 16, and Anthropic made Claude Sonnet 5's $2/$10 rates permanent.

FAQ#

Which is the cheapest AI coding model in October 2026?#

GPT-6 Luna and Claude Haiku 5.5 tie at $0.10 per million input tokens and $0.50 per million output, with $0.01 cache reads (verified October 8, 2026). Luna keeps that rate up to 272K input tokens; Haiku 5.5 keeps it only up to 100K. DeepSeek V4.1-Flash (deepseek-flash) at $0.15/$0.60 off-peak is the open-weights floor, and its rates double during weekday peak hours.

Is Claude Haiku 5.5 cheaper than GPT-6 Luna?#

Only on long prompts does the answer turn clearly against Haiku. They tie on list price for prompts up to 100K tokens. Above 100K, Haiku 5.5 bills $0.50/$2.50 against Luna's $0.10/$0.50 (Luna only rises to $0.20/$0.75 past 272K), so our 200K-in, 20K-out example costs $0.030 on Luna and $0.15 on Haiku 5.5 at the same token count. Claude's tokenizer also counts about 30% more tokens than Haiku 4.5 did, so measure real prompts on both APIs. The tie is on price only: Anthropic's vendor-reported benchmarks put Haiku 5.5 ahead of Luna on agentic coding and computer use.

Is DeepSeek V4.1-Flash good at agent tasks?#

DeepSeek's September 10 release notes report third-party tests putting V4.1-Flash ahead of V4-Pro on performance, cost, and speed, and its predecessor (V4-Flash 0731) posted the strongest agent benchmark claims ever published for the budget tier (82.7 Terminal Bench 2.1, 54.4 DeepSWE, vendor-reported). The V4.1 benchmark tables ship as images in the release news; treat them as vendor-reported until independent evals land.

What is the difference between GPT-6 Luna and DeepSeek V4.1-Flash?#

Both are sub-$0.20 input models. GPT-6 Luna (proprietary) brings OpenAI's tool-calling and Responses API stack, a 1,050,000-token context window, and costs $0.10/$0.50. V4.1-Flash (MIT open weights) costs $0.15/$0.60 off-peak, can be self-hosted, and adds native visual understanding.

Should I still use Claude Haiku 4.5?#

Only while you migrate. Haiku 4.5 lists at $1/$5, and Anthropic prices Haiku 5.5 90% lower for requests up to 100K tokens. The switch is not just a model id change - fixed thinking budgets, sampling parameters and assistant prefill return 400 errors on Haiku 5.5 - so follow the migration steps before moving production traffic.

Is Gemini 3.5 Flash worth the higher price?#

On price alone, rarely: at $1.50/$9.00 it costs 15x Luna and Haiku 5.5 on input, and Google's own Gemini 3.8 Flash lists at $0.75/$3.75 through December 31, 2026. It makes sense when the rest of your stack is already on Google and your golden-set completion rate on Gemini is clearly higher.

Should I route between budget models and frontier models?#

Yes, for volume. The standard split is budget models for high-volume, well-understood work and frontier models for the hard tail. Both ends moved this autumn: GPT-6 Luna and Haiku 5.5 handle short tool steps at ten cents per million input tokens, and GPT-6.1 Sol at $2/$10 keeps the frontier's entry price near what used to be mid-tier.

Official Sources#

Sources#

• OpenAI: API Pricing - accessed October 8, 2026
• OpenAI: GPT-6 Luna model page - accessed October 8, 2026
• OpenAI: GPT-5.6 Luna model page - accessed October 8, 2026
• OpenAI: Deprecations - accessed October 8, 2026
• Anthropic: Claude Haiku 5.5 overview - accessed October 8, 2026
• Anthropic: Introducing Claude Haiku 5.5 - announced October 7, 2026, accessed October 8, 2026
• Anthropic: Claude Haiku 5.5 migration guide - accessed October 8, 2026
• DeepSeek: Models & Pricing - accessed October 8, 2026
• DeepSeek: V4.1-Flash Release News - accessed September 26, 2026
• Google: Gemini API Pricing - accessed October 8, 2026

Continue Reading#

• Claude Haiku 5.5: Pricing, Migration Changes, and When to Use It - the 100K cliff, the tokenizer catch, and the ten migration steps from Haiku 4.5
• Claude Max API Credits: $100 a Month, What They Cover - the monthly API credit that stretches furthest on Haiku 5.5
• Frontier Model API Pricing - the full provider landscape with same-day-verified prices
• GPT-6 Sol vs Claude Opus 5.5 vs Grok 4.7: The September 2026 Price War - the frontier launches and what they mean for the tier below
• Model Routing Strategies for Cost-Effective Coding - when to route between tiers and when routing does not help
• OpenAI Cuts GPT-5.6 Luna by 80% - the July 30 price reset, with the HN reaction
• Where to Access AI Models - official APIs, routers, and free tiers for every model in this comparison
• Best Open-Source Models - the open-weight options when list price is the deciding factor