The Mid-Tier Shootout: GPT-5.4 vs Gemini 3.1 Pro vs DeepSeek V4 Pro

Last updated: July 31, 2026

Flagship comparisons get the headlines. Fable 5 at $10/$50, GPT-5.5-pro at $30/$180 - those are the models people argue about on launch day. But most production token spend runs a tier below: the $2-3 per million input token class that handles summarization queues, agent steps, RAG answers, and the long tail of app traffic that never needed a frontier model.

What Changed on July 31, 2026#

• GPT-5.6 Terra entered the tier at exactly $2/$12 (verified July 31, 2026 on the OpenAI pricing page). It is the same price as Gemini 3.1 Pro under 200K context, with the same 1M context ceiling and a 50% batch discount. GPT-5.4 remains listed at $2.50/$15 but the family has been superseded - new integrations should start with Terra. Our GPT-5.6 Sol/Terra/Luna guide covers the family.
• Gemini 3.1 Pro is still in preview as of July 31, 2026 - now five and a half months since its February release. Pricing is unchanged at $2/$12 under 200K context.
• DeepSeek V4 prices unchanged ($0.435/$0.87 for V4 Pro), but the July 24 deprecation wave retired the legacy deepseek-chat/deepseek-reasoner names in favor of deepseek-v4-pro/deepseek-v4-flash, and the V4 Flash 0731 update shipped. A peak/off-peak pricing policy is announced but not yet effective - peak-hour rates will be 2x.
• GPT-5.6 Luna makes the tier below this one more interesting - at $0.20/$1.20 it overlaps with DeepSeek V4 Flash pricing. See our budget coding models comparison.

That tier has three serious contenders right now: OpenAI's GPT-5.4, Google's Gemini 3.1 Pro Preview, and DeepSeek's V4 Pro. They land within a few points of each other on the benchmarks that matter, so the decision comes down to pricing mechanics, context behavior, and license terms - the boring details that compound into real money at volume. Every price below is verified against the live vendor pages.

The Contenders at a Glance#

All prices verified July 31, 2026 against the OpenAI pricing page, the Gemini API pricing page, and the DeepSeek pricing page. Benchmark figures come from third-party writeups (sources at the end) because vendor benchmark reporting is inconsistent across variants and scaffolds - treat the coding scores as a cluster, not a ranking.

Note for new builds: OpenAI's GPT-5.6 Terra matches Gemini 3.1 Pro's $2/$12 head-on and is the natural GPT-5.4 successor in this tier, so consider it as a fourth contender if you are evaluating today. The table above keeps GPT-5.4 because it remains listed and priced on OpenAI's page.

Pricing: Where the Tier Stops Being a Tier#

The headline framing of "the $2-3 class" undersells how wide the spread actually is once you do the math.

GPT-5.4 runs $2.50 input / $15.00 output, with cached input at $0.25 and a 50% batch discount (verified July 31, 2026, developers.openai.com/api/docs/pricing). Its successor GPT-5.6 Terra runs $2.00 / $12.00 with cached input at $0.20 - the new best price from OpenAI in this tier. Gemini 3.1 Pro Preview is $2.00 / $12.00 for prompts up to 200K tokens, jumping to $4.00 / $18.00 above that, with batch at half price (verified July 31, 2026, ai.google.dev/gemini-api/docs/pricing).

Then there is DeepSeek V4 Pro: $0.435 input on a cache miss, $0.87 output, and a cache-hit input price of $0.003625 per million tokens (verified July 31, 2026, api-docs.deepseek.com/quick_start/pricing). That output price is roughly 17x cheaper than GPT-5.4 and about 14x cheaper than Gemini 3.1 Pro under the 200K threshold.

One thing worth flagging: April launch coverage, including Artificial Analysis and DataCamp, cited V4 Pro at $1.74 / $3.48. The official page as verified July 31, 2026 lists $0.435 / $0.87. The live page is the source of truth - and it says V4 Pro costs about a quarter of what most comparison articles still quote.

Concrete example: a workload pushing 10M input and 2M output tokens per month, no caching, all prompts under 200K:

• GPT-5.4: $25.00 + $30.00 = $55.00
• Gemini 3.1 Pro Preview: $20.00 + $24.00 = $44.00
• DeepSeek V4 Pro: $4.35 + $1.74 = $6.09

GPT-5.4 and Gemini are within 25% of each other. DeepSeek is playing a different sport. If raw unit cost is your primary axis, this comparison is over before the benchmarks section - which is exactly why the benchmarks section matters.

Caching Mechanics Differ More Than the Sticker Prices#

All three discount repeated input, but the mechanics diverge. OpenAI's cached input is a flat $0.25 (0.1x base) with no storage fee. Gemini charges $0.20 per million for cache reads under 200K context plus $4.50 per million tokens per hour of cache storage, so idle caches cost money on spiky traffic. DeepSeek's cache hits are $0.003625 - close enough to free that input-side prompt architecture barely matters for cost. If your agent reuses a large system prompt thousands of times a day, these mechanics can outweigh the base rates. For the routing-layer view, see our LLM router comparison.

Benchmarks: Close Enough That Price Is the Tiebreaker#

On agentic coding, the three are clustered. The nxcode GPT-5.4 guide puts GPT-5.4 at 57.7% on SWE-bench Pro and roughly 80% on SWE-bench Verified. The nxcode Gemini 3.1 Pro guide lists Gemini 3.1 Pro at 54.2% on SWE-bench Pro (Public) and 80.6% on SWE-bench Verified. DataCamp's DeepSeek V4 analysis has V4 Pro at 55.4% on SWE-bench Pro and 67.9% on Terminal-Bench 2.0, versus Gemini's 68.5% on the same benchmark.

A 3.5-point spread reported by different evaluators with different scaffolds is not a ranking - it says all three do real agentic coding at a similar hit rate. The differentiation shows up off the center line:

• GPT-5.4 reports 75% on OSWorld, which nxcode notes is above the 72.4% human expert baseline for desktop automation. If your workload involves computer use, GPT-5.4 has the strongest claim in this tier.
• Gemini 3.1 Pro posted a verified 77.1% on ARC-AGI-2, more than double Gemini 3 Pro, per Google's announcement, plus 94.3% on GPQA Diamond per nxcode. For novel-reasoning and science-heavy workloads, Gemini has the edge.
• DeepSeek V4 Pro scores an Artificial Analysis Intelligence Index of 52, the #2 open-weights model, with a GDPval-AA of 1554 leading the open-weights field, per Artificial Analysis. DataCamp's framing is the honest one: it trails closed state of the art by 3 to 6 months, at a fraction of the price.

One caveat from the Artificial Analysis data: V4 Pro burned 190M output tokens completing the Intelligence Index. DeepSeek thinks in volume, so part of the per-token advantage gets spent on extra tokens - V4 Pro still wins on total cost, but do not expect the full 17x in practice.

Context Windows and Long-Context Behavior#

All three advertise 1M-token context, but the fine print differs in ways that bite.

GPT-5.4 gives you 272K standard, 1M via API - and input doubles to $5.00 per million above the 272K threshold per the nxcode guide; OpenAI's pricing page splits the model into short-context and long-context rows. Gemini 3.1 Pro tiers earlier: above 200K prompt tokens, input goes to $4.00 and output to $18.00 (verified July 31, 2026 on the Gemini pricing page). DeepSeek V4 Pro lists no long-context surcharge at all, plus the largest max output of the three at 384K tokens (verified July 31, 2026).

The asymmetry to remember: Gemini has the smallest max output (64K) and the earliest price cliff (200K), GPT-5.4 has the smallest standard context (272K), and DeepSeek is the only one where a 900K-token prompt costs the same per token as a 9K one. DataCamp reports V4 Pro at 83.5% on the MRCR 1M needle test, so the long context is not just nominal.

One operational note from DeepSeek's pricing page: V4 Pro is capped at 500 concurrent requests (V4 Flash gets 2,500). For high-fanout workloads, that ceiling is a real constraint the closed vendors do not impose at this tier.

License and Deployment Terms#

This is the dimension where the three models genuinely fork.

DeepSeek V4 is MIT-licensed with downloadable weights - 1.6T total parameters, 49B active, an 865GB download, per DataCamp. You can self-host it, fine-tune it, and ship it in commercial products with no usage-based terms. The practical bar is high, but for teams with data-locality requirements it is the only option in this tier that fully removes the API dependency. We covered the economics in our DeepSeek V4 developer guide.

GPT-5.4 is proprietary but GA, with the most mature surrounding tooling, and OpenAI offers regional data-residency processing at a 10% uplift for models released on or after March 5, 2026 (verified July 31, 2026 on the pricing page). For setup details see our GPT-5.4 developer guide.

Gemini 3.1 Pro is the odd one out on status: five and a half months after its February 19 release, the model string is still gemini-3.1-pro-preview (verified July 31, 2026 on the pricing page), and Google's announcement promises GA "soon" without a date. Preview means Google reserves the right to change behavior and endpoints. Plenty of teams run preview models in production anyway, but if your change management cares about stability guarantees, the label matters. To evaluate it cheaply from the terminal first, our Gemini CLI guide covers the free-tier path.

Decision Guide by Persona#

• Cost-driven, high-volume production traffic. DeepSeek V4 Pro, and it is not close. At $0.435 / $0.87 with near-free cache hits, you can run roughly 9x the volume of GPT-5.4 on the same budget. Budget for its verbose reasoning style and the 500-request concurrency cap. Our budget coding agents writeup has real-world patterns.
• Computer-use and desktop automation agents. GPT-5.4 (or GPT-5.6 Terra for new builds). The 75% OSWorld figure is the standout benchmark in this tier, and GA status reduces integration risk.
• Long-document analysis under 200K tokens. Gemini 3.1 Pro Preview. At $2.00 / $12.00 it is the cheapest closed-vendor option alongside GPT-5.6 Terra, and its reasoning scores lead the tier. Architect around the 200K price cliff and 64K output ceiling.
• Compliance, data locality, or vendor independence. DeepSeek V4 Pro via self-hosted MIT weights, or GPT-5.4 with regional processing if self-hosting is off the table.
• Preview-averse enterprises. GPT-5.4, GPT-5.6 Terra, or DeepSeek V4 Pro. Gemini 3.1 Pro is the only one of the three still labeled preview as of July 31, 2026 - GPT-5.4 is GA, and DeepSeek's API docs list deepseek-v4-pro with no preview label (the DeepSeek pricing page retired older model names in its favor on July 24, 2026).

When to Skip This Tier#

Skip up if your workload is long-horizon agentic work where completion rate dominates cost. A mid-tier model that fails a complex migration three times costs more than a frontier model that finishes once - the same math we walked through in Fable 5 vs DeepSeek V4 on cost versus quality. The 55-58% SWE-bench Pro cluster here converts directly into retries on genuinely hard autonomous tasks.

Skip down if your workload is classification, extraction, or short-form generation at scale. GPT-5.6 Luna is $0.20 / $1.20 and DeepSeek V4 Flash is $0.14 / $0.28 (both verified July 31, 2026 on their pricing pages). For tasks that do not need mid-tier reasoning, even V4 Pro is overpaying.

Stay in this tier for everything in between, which for most products is the majority of tokens. The honest summary: GPT-5.4 buys polish and GA stability, Gemini 3.1 Pro buys reasoning headroom and the best closed-vendor price under 200K context, and DeepSeek V4 Pro buys an order of magnitude on cost in exchange for a 500-request concurrency cap, verbose reasoning-token usage, and a few months of capability lag.

FAQ#

Is DeepSeek V4 Pro really cheaper than GPT-5.4 and Gemini 3.1 Pro?#

Yes, by a wide margin on list price. Verified July 31, 2026 on the official pricing pages: V4 Pro is $0.435 input / $0.87 output per million tokens versus $2.50 / $15.00 for GPT-5.4 and $2.00 / $12.00 for Gemini 3.1 Pro under 200K context. DeepSeek emits more reasoning tokens per task, so realized savings are smaller than the sticker ratio, but still large.

Which model is best for coding in the mid-tier class?#

They are close: GPT-5.4 reports 57.7% on SWE-bench Pro, DeepSeek V4 Pro 55.4%, and Gemini 3.1 Pro 54.2% (Public split), per the third-party sources below. Different evaluators produced these numbers, so treat them as a cluster and pick on price, context behavior, and tooling instead. New builds should also compare GPT-5.6 Terra, which matches Gemini's $2/$12 and is OpenAI's current mid-tier offering.

Is Gemini 3.1 Pro still in preview?#

Yes. As of July 31, 2026, the Gemini API pricing page still lists the model as gemini-3.1-pro-preview, and Google's announcement promises general availability "soon" without a date. It has been in preview since February 19, 2026 - over five months.

Can I self-host DeepSeek V4 Pro?#

Yes. The weights are MIT-licensed, but V4 Pro is a 1.6T-parameter mixture-of-experts model with an 865GB download, so self-hosting needs serious multi-GPU infrastructure. The smaller V4 Flash (284B total, 160GB) is the more realistic self-host target.

Does long context cost extra on these models?#

On two of the three. Gemini 3.1 Pro doubles input to $4.00 and raises output to $18.00 above 200K prompt tokens, and GPT-5.4 input doubles to $5.00 above its 272K standard window per third-party documentation. DeepSeek V4 Pro lists flat pricing across its full 1M context (all verified July 31, 2026).

Official Sources#

Sources#

• OpenAI API pricing - GPT-5.4 and GPT-5.6 family pricing, caching, batch, regional uplift (accessed July 31, 2026)
• Gemini API pricing - Gemini 3.1 Pro Preview tiers, caching, batch (accessed July 31, 2026)
• DeepSeek API pricing - V4 Pro and V4 Flash pricing, context, concurrency, deprecations (accessed July 31, 2026)
• Google: Gemini 3.1 Pro announcement - release date, ARC-AGI-2, availability (accessed July 31, 2026)
• nxcode: GPT-5.4 complete guide - GPT-5.4 benchmarks, variants, context tiers (accessed July 31, 2026)
• nxcode: Gemini 3.1 Pro complete guide - Gemini 3.1 Pro benchmark set (accessed July 31, 2026)
• DataCamp: DeepSeek V4 - V4 specs, benchmarks, positioning (accessed July 31, 2026)
• Artificial Analysis: DeepSeek V4 Pro and V4 Flash - Intelligence Index, GDPval-AA, token usage (accessed July 31, 2026)

Continue Reading#

• GPT-5.6 Sol/Terra/Luna Guide - OpenAI's current mid-tier lineup and where Terra fits
• Budget AI Coding Models Compared July 2026 - the tier below this one, including Luna and V4 Flash 0731
• DeepSeek V4 Economics - how V4's cost advantage holds up in agentic workloads
• Model Routing Strategies for Cost-Effective Coding - routing across tiers instead of picking one model
• DeepSeek Chat to V4 Migration Guide - what the July 24 legacy-name retirement meant for API callers