Budget AI Coding Models Compared July 2026: V4 Flash vs Luna vs Gemini 3.5 Flash vs Haiku 4.5

Last updated: July 31, 2026

Two events in the last 48 hours make the cheap tier of coding models impossible to ignore. On July 30, OpenAI cut GPT-5.6 Luna by 80% to $0.20 per million input tokens. On July 31, DeepSeek shipped V4 Flash 0731 - a re-post-trained version of its budget model that posts agent benchmark scores well past what V4 Pro Preview managed in the spring, at the same $0.14/$0.28 prices. A tier that used to mean "good enough for extraction and classification" is now asking to handle whole agent loops.

This comparison covers the four models that anchor that tier in July 2026: DeepSeek V4 Flash, GPT-5.6 Luna, Gemini 3.5 Flash, and Claude Haiku 4.5. Every price below was verified against the live vendor pricing pages on July 31, 2026. Benchmark coverage is honest about what is third-party-verified and what is vendor-reported - the two are not the same thing.

The Quick Comparison#

All prices verified July 31, 2026: DeepSeek pricing, OpenAI pricing, Gemini pricing, Anthropic pricing.

The price spread across the tier is now wider than the capability spread on many tasks. V4 Flash and Luna are in a class of their own on cost - 5 to 10x cheaper than the two hosted options on input, and the cache-hit rates are close to free. The interesting question is no longer "which is cheapest" but "how much agentic capability does each model actually bring to a loop."

DeepSeek V4 Flash 0731: The Re-Post-Trained Agent Workhorse#

DeepSeek's official change log describes the July 31 update precisely: "keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained." Same 1M context, same 384K max output, same $0.14/$0.28 pricing, same deepseek-v4-flash model name - but the benchmark table reads like a different model.

The vendor-reported numbers from the pricing page: Terminal Bench 2.1 at 82.7, DeepSWE at 54.4, Cybergym at 76.7. Those are long-horizon agent benchmarks - real terminal sessions and issue-resolving loops, not multiple choice - and they are the strongest claim this tier has ever made. Two caveats before you take them at face value: they are DeepSeek's own runs (the DeepSeek Harness is "to be released soon"), and the two biggest developer-facing numbers, Terminal Bench and DeepSWE, are the ones to benchmark against other models yourself.

What is not a caveat is the access story. V4 Flash is the only DeepSeek model with native Responses API support today, and DeepSeek publishes a one-line Codex setup script that wires the model into Codex CLI, the ChatGPT desktop app, and the VS Code extension. The economics post we wrote when V4 launched pinned Flash as an inner-loop model; the July 31 update analysis walks through what changed when that model started handling whole loops.

GPT-5.6 Luna: The 80% Cut Changes the Routing Math#

OpenAI's July 30 announcement cut Luna from $1/$6 to $0.20/$1.20 - an 80% drop - and Terra by 20% to $2/$12. Luna's cache hit is $0.02 per MTok, and long-context rates run $0.40/$1.80. Our breakdown of the announcement covered the reaction: a typical agent conversation of 50K input and 10K output tokens now costs about $0.022, cheap enough that the abstraction overhead of a router may not justify itself for high-volume work.

Luna's position in the tier is interesting because it is the only proprietary model here with a first-party agent stack behind it - Codex, Responses API, structured outputs, and the same tool-calling machinery as Sol. The HN thread around the cut included the counterpoint that Luna is "still more expensive than DeepSeek V4 Flash" per token, and the reply was the real argument: Luna handles tool calls and multi-step workflows that cheaper models cannot reliably execute. That capability gap is exactly what the 0731 update challenges on the DeepSeek side.

One note on naming: Luna is the entry tier of the GPT-5.6 family, and OpenAI's deprecation clock now applies to the original GPT-5 snapshots, not to Luna - it is the recommended replacement for gpt-5-nano and gpt-5-mini on the deprecations page.

Gemini 3.5 Flash and Claude Haiku 4.5: The Hosted Middle#

Google and Anthropic do not compete on price in this tier, and they do not need to - their budget models win on integration quality and provider-level guarantees.

Gemini 3.5 Flash at $1.50/$9.00 (verified July 31) has been GA since May and is the only model in this tier with first-party grounding with Google Search built in - 5,000 free search requests per month across Gemini 3.x models, then $14 per 1,000. If your workload needs fresh web context in the loop, that changes the cost math entirely: a search-grounded flash model at $1.50 input competes on total cost with cheaper models plus a separate search bill. Cache hits run $0.15 with $1 per MTok per hour storage. 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.

Claude Haiku 4.5 at $1/$5 (verified July 31) is the middle option on price and the strongest on Anthropic-side integration: it is the default model for Claude Code's fast paths, inherits the 1M context window at flat pricing, and slots into any Anthropic API workflow with zero migration. Its role in the tier is the safety pick - if the rest of your stack is Claude, keeping the budget model in the same provider removes the wire-format and eval-surface differences that cross-provider routing introduces.

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:

• DeepSeek V4 Flash: $0.028 + $0.0056 = about $0.034
• GPT-5.6 Luna: $0.040 + $0.024 = $0.064
• Claude Haiku 4.5: $0.20 + $0.10 = $0.30
• Gemini 3.5 Flash: $0.30 + $0.18 = $0.48

So the spread is about 14x across the tier for the same token count. The catch is that token counts are not the same: 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 a model that needs five attempts at $0.034. 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: a peak/off-peak pricing policy is coming, with 2x prices during 9:00-12:00 and 14:00-18:00 Beijing time. For batch-able work, that makes off-peak scheduling real money - batching agent work into the cheap window is how the already-low numbers get lower.

Benchmarks: What Is Verified, What Is Not#

That table is thin, and that is the honest state of the tier. DeepSeek publishes hard agent numbers (vendor-reported, with config documented). The other three vendors publish marketing-tier claims and point to their flagship models for the benchmark pages. If you are choosing among these four, the benchmark evidence is currently a DeepSeek-first story - which is exactly why independent evals matter more here than at the frontier, where multiple third parties benchmark every release.

What the tier does have in common: all four models are cheap enough that the cost of running your own golden-set evaluation is trivial. Our eval-before-switch day plan applies verbatim - build a golden set, run it against all four, 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: DeepSeek V4 Flash or GPT-5.6 Luna. Either is under a cent per task; pick by whether you want open weights (DeepSeek, MIT license) or first-party tool calling (Luna).
• Agent loops with tools, moderate volume: GPT-5.6 Luna if you are on the OpenAI stack, Claude Haiku 4.5 if you are on Anthropic. The wire compatibility and the provider's agent tooling matter more than the token price.
• Search-grounded or retrieval-heavy work: Gemini 3.5 Flash. The built-in grounding at $14 per 1,000 searches past the free tier changes the total-cost comparison for anything needing live web context.
• Self-hosted or data-sensitive: DeepSeek V4 Flash. MIT open weights, 1M context, and the only model in this tier you can run on your own hardware.
• Everything else: measure. The 14x 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 - Opus 5 at $5/$25 and GPT-5.6 Sol at $5/$30 are the current defaults, and a failed hour-long run costs more in developer time than the tier difference saves in tokens. Second, output-heavy reasoning: Luna's $1.20 output and Flash's $0.28 output are still orders of magnitude below the frontier's $25-$30, 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. The model routing strategies post covers when routing between tiers pays and when it does not.

FAQ#

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

DeepSeek V4 Flash at $0.14 per million input and $0.28 per million output tokens, verified July 31, 2026. GPT-5.6 Luna is the closest competitor at $0.20/$1.20 after its July 30 price cut.

Is DeepSeek V4 Flash 0731 good at agent tasks?#

DeepSeek's own numbers (not yet third-party verified) put it at 82.7 on Terminal Bench 2.1 and 54.4 on DeepSWE - the strongest agent benchmark claims ever published for the budget tier, from a model that costs $0.14/$0.28 and supports the Responses API and Codex integration.

What is the difference between GPT-5.6 Luna and DeepSeek V4 Flash?#

Both are sub-$0.50 input models. Luna (proprietary) brings OpenAI's tool-calling and Responses API stack and costs $0.20/$1.20. V4 Flash (MIT open weights) costs $0.14/$0.28, can be self-hosted, and just shipped a re-post-training update aimed at agent workloads.

Is Gemini 3.5 Flash or Claude Haiku 4.5 worth the higher price?#

For search-grounded work, Gemini 3.5 Flash's built-in grounding can make it cheaper in total cost than a cheaper model plus a separate search bill. Claude Haiku 4.5 is the value pick when the rest of your stack is Anthropic - same provider, same wire format, 1M context at flat pricing.

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. The July 2026 updates narrow the gap on the volume side - V4 Flash 0731 and Luna at $0.20 both handle tool loops that used to require a step up.

Official Sources#

Sources#

• DeepSeek: Models & Pricing - accessed July 31, 2026
• DeepSeek: Change Log - accessed July 31, 2026
• OpenAI: API Pricing - accessed July 31, 2026
• OpenAI: Model Deprecations - accessed July 31, 2026
• Google: Gemini API Pricing - accessed July 31, 2026
• Anthropic: API Pricing - accessed July 31, 2026

Continue Reading#

• DeepSeek V4 Flash 0731: The Budget Tier Just Overtook Pro Preview - what the re-post-training actually changed
• OpenAI Cuts GPT-5.6 Luna by 80% - the July 30 price reset, with the HN reaction
• Frontier Model API Pricing: July 2026 - the full provider landscape with same-day-verified prices
• Model Routing Strategies for Cost-Effective Coding - when to route between tiers and when routing does not help
• GPT-5.6 Sol, Terra, and Luna: What Each Model Is For - the OpenAI tier ladder above this comparison