How to Choose DeepSeek, GPT, and Gemini on OpenRouter: Coding, Writing, and Reasoning

Pin three slugs before you compare brands

OpenRouter’s product is an OpenAI-compatible key that can point at many models. The failure mode is comparing names. DeepSeek is not only V3.2; a higher-compute Speciale sits beside it. The GPT line still lists 5.2 next to 5.4. Gemini still lists 2.5 Pro, 3 Pro Preview, and 3.1 Pro Preview. A brand ranking treats different prices, context windows, and modalities as one SKU. The first configuration change is the slug, not the slogan.

The three pinned here are the variants whose public OpenRouter pages, in October 2026, still show a checkable price and context window for each product line. DeepSeek V3.2 shipped on 1 December 2025, with a 163,840-token window. The page says you can turn reasoning on with reasoning.enabled, and it describes gold-medal-class results on the 2025 IMO and IOI plus post-training aimed at tool use. GPT-5.4 is described as the frontier model that folds the Codex and GPT lines together: about 1.05 million tokens of context (roughly 922K input and a 128K output cap) and text plus image input. Gemini 3.1 Pro Preview is still a preview: 1,048,576 tokens of context, a 65,536-token output cap, and text, image, video, audio, and code. Preview means the behavior can move. Production needs a second slug behind it.

Two posts on this site answer different questions. 2026 OpenRouter Top 10 ranks a crowd of models; it does not assign these three slugs to jobs. If the invoice is already painful enough that you want to leave OpenRouter, use the OmniRoute migration checklist. This article stays on OpenRouter and only answers where the default request should go for coding, writing, and reasoning.

The pages to re-open before you spend are DeepSeek V3.2, GPT-5.4, and Gemini 3.1 Pro Preview. You still point base_url at OpenRouter and change the model string. The quickstart is enough to wire that. Every price and score below carries the date it was read. Provider discounts, cache hits, and routing mode will make your invoice differ from the table. Treat the live model page as the quote.

One key, three bills

These are not three sticker prices. Each call is model times provider times route. In October 2026 the DeepSeek comparison page headlined about $0.2088 per million input tokens and $0.3096 per million output tokens. That headline is a discounted provider. On the same model page, providers run from roughly $0.26 / $0.38 up to $3 / $4.50. A SiliconFlow row in mid-October still showed about $0.259 in and $0.80 out. Cache reads are separate; the cheap end is about $0.0216 per million. If you do not pin a provider, Balanced routing does not promise the headline.

GPT-5.4’s standard route is easier to quote: $2.50 in, $15 out, $0.25 cache read, per million tokens. Flex (about $1.25 / $7.50) and Fast (about $5 / $30) are listed and are not part of standard routing. Gemini 3.1 Pro Preview headlines at $2 in and $12 out, with cache reads around $0.20. Web search is extra; the model page cites about $14 per thousand calls. Output caps differ too: about 128K for GPT-5.4 and 65,536 for Gemini 3.1. A long draft fails the cap before it fails the unit price.

One mix that is easy to recompute: 100,000 input tokens and 8,000 output tokens. At the comparison-page headline, DeepSeek is about $0.023. On the $0.259 / $0.80 row it is about $0.032. The same call is about $0.37 on GPT-5.4 and about $0.30 on Gemini 3.1. In that mix GPT is roughly 16× the headline DeepSeek price and about 11× the $0.80-output row. Fatter outputs widen the gap, because $15 versus $0.31 is not the same ratio as $15 versus $0.80. Cache hits discount the input side later. None of the three skip the full price on a cold first call.

OpenRouter names the routes Balanced (price plus speed), Nitro (faster), Floor (cheaper), and Exacto (higher tool-call accuracy). If an agent loses because the tool arguments are wrong, try Exacto before you blame the slug. If a writing batch is only a first draft, Floor plus V3.2 usually beats “always GPT-5.4” on the invoice. Latency is not an SLA you can paste from a blog. The 5 October 2026 comparison snapshot showed GPT-5.4 against Gemini 3.1 at about 1.74s versus 90.9s median latency and about 42 versus 8 tokens per second. A 90-second median is a slice of a reasoning endpoint or a cold provider, not a law that Gemini is always fifty times slower. Read the live activity chart for the route you actually selected before you hard-code a timeout.

Coding: drafts, repos, and a terminal

Coding is not one score. A one-file patch, a cross-file API change, and a terminal session that has to turn tests green can crown three different models. Vector Wire’s 8 October 2026 shared-test writeup of DeepSeek-V3.2 versus GPT-5.4 is useful because the gaps are uneven: SWE-bench Pro about 15.6 versus 57.7, Terminal-Bench 2.1 about 46.8 versus 78.3, and SWE-bench Multilingual only about 1.5 points apart (70.2 versus 71.7). Short, multilingual, single-point fixes are a fair job for V3.2. Long, multi-file work that has to run commands in a real terminal is where the published GPT-5.4 scores pull away. Those are aggregated public scores, not your repo’s pass rate. Change the scaffold and the order moves.

OpenRouter’s own 5 October 2026 comparison gives DeepSeek V3.2 (Reasoning) a Coding Index of 44.2 and an Intelligence Index of 21.5. The other column on that page is GPT-5.2, not 5.4; GPT-5.2’s Intelligence Index there is 30.4 and its Coding Index was blank. A different page, GPT-5.4 versus Gemini 3.1 Pro Preview, lists 39.0 and a Coding Index of 71.1 for GPT-5.4, with both Gemini indexes empty that day. You cannot staple “71 versus 44” into an official GPT-versus-DeepSeek verdict. The variants and the opponents differ. The narrower claim is the one you can put in config: the coding index OpenRouter actually publishes is, for now, filled in on GPT-5.4; V3.2 has its own coding index against a different opponent.

Day to day, three lanes are enough. Comments, a single unit test, a Chinese or English explanation of a stack trace, and a minimal diff are V3.2 work: ten tries often still cost less than one GPT call. If the change has to touch callers, types, and a migration script together, or the agent must install a dependency, read the failure log, and edit again, switch the default to openai/gpt-5.4. Past roughly 160K input tokens, V3.2’s window is full. GPT-5.4 and Gemini 3.1 are both around a million. For pure code and text, GPT-5.4’s terminal and tool record is the closer fit. Video of a UI flow, a design-review recording, or a voice note is the reason to call Gemini—not “who has the longer window.” Those two windows are already the same class.

GPT-5.4’s page is explicit about image input. Gemini 3.1 also takes video and audio. A single error screenshot is fair game for both. A screen recording that the model must watch before it edits code is, among these three, a Gemini job. When the preview returns empty, or a tool turn loses its place, send the same prompt to GPT-5.4. Do not let the gateway retry the same preview forever.

Writing: volume, finished English, long sources

Writing has no stable public board that all three vendors publish at the level of SWE-bench. A single prose ELO will be stale next month. Split the work into the four jobs you actually pay for: a high-volume first draft, finished English, a format with hard constraints, and a draft that must stay attached to a long source and must not invent citations.

Internal notes, support macros, and a first Chinese or English pass are usually a V3.2 default. The call is cheap enough to generate three lengths and let a person pick. When a style guide already exists, a release note is banned from marketing voice, or sections must follow the headings you supplied, the published gap that matters is instruction following, not “who writes prettier sentences.” Vector Wire’s October 2026 scoreboard puts instruction following among the areas GPT-5.4 wins outright against V3.2. That is still their prompt set, not your banned-word list. Regress against ten of your own old posts before you trust another roundup.

Long sources are a separate failure. Meeting notes, a paper packet, or a regulatory extract past 160K tokens will be truncated on V3.2. The summary still looks complete. It never saw the second half. GPT-5.4 and Gemini 3.1 can both hold on the order of a million tokens. Gemini’s extra reasons are the medium—video or audio—and an output price of $12 versus $15. If the failure you hate is a page number that does not exist, both models need the prompt to quote the source span, and a script should check the quote. A more expensive slug does not cancel fabrication.

Score writing on four columns, 0 to 2 each: constraints hit (length, person, banned words), fabrication (links, numbers, quotations that are not in the source), revision rounds, and dollars of output tokens. Same temperature, same brief, one run each. The Chinese or internal-draft column often lands on V3.2 because a second round is still cheaper than GPT’s first round. The external English column often lands on GPT-5.4 because one avoided brand rewrite pays for the unit price. A brief that includes a recording is what puts Gemini in the final. Do not average the columns into a house champion. The average mixes a cheap draft with an expensive finished piece and calls the mush a winner.

Reasoning: math, science, multi-step tools

Reasoning splits the same way. Contest-math volume, graduate science items, and multi-step tool use do not share an objective. DeepSeek’s model copy talks about sparse attention, scaled reinforcement-learning post-training, and gold-medal-class 2025 IMO and IOI results. That is the right story when you have a large set of auto-graded math items and you want the cost per item under a few cents. It is not the sentence “first on every shared board.” Vector Wire’s 8 October 2026 comparison says GPT-5.4 leads on the reasoning, math, and factual tests both models publish: GPQA Diamond about 92 versus 84, Humanity’s Last Exam about 43.7 versus 24.6, and a much wider ARC-AGI-2 gap (about 73.3 versus 4). Sweep with V3.2. When a single solution must satisfy a picky rubric, call GPT-5.4.

Science items need two rulers kept apart. OpenRouter’s 5 October 2026 page gives GPT-5.4 an Intelligence Index of 39.0 and Gemini 3.1 Pro Preview 28.0, with Gemini’s coding index empty. Other writeups from spring 2026 quote higher GPQA and SWE-bench Verified numbers for Gemini 3.1 Pro. Both can be true because the item set, the reasoning effort, and the scaffold differ. If you ship biomedical or regulatory answers, run your own 20 checked questions and mark whether the miss was a fact or a step. Do not use a 94 on board A to erase a 28 on board B, or the reverse.

Multi-step tools are where the Gemini preview is easiest to break and where the docs are the most specific. OpenRouter tells you to preserve reasoning blocks across tool turns. The page is the reasoning-token guide. If your gateway returns only the visible text and drops reasoning details, the second turn looks like amnesia. GPT-5.4’s product story puts the coding agent and the general model in one system, so the tool loop is the path the page is written around. V3.2 uses reasoning.enabled to decide whether this call spends reasoning tokens. When all three “support tools,” check that your client sends the reasoning block back unchanged before you compare IQ. A large share of “Gemini will not follow the plan” tickets are client bugs.

How to choose: a table and one route

The table below is meant to sit next to the gateway. Start on the cheap row. Escalate only when the window, the modality, or the cost of another revision has already beaten the unit price.

The snippet is a starting policy. Change the thresholds after your own 20 tasks. Those tasks do not need to be public, but they do need to be your real failures: a unit test, a cross-file change, a terminal log, an internal note, a release note, a banned word, a fake page number inside a long source, a second tool turn that must remember the first constraint, a screenshot, a recording. Score 0 to 2 per task, by column. Do not collapse them into one total.

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek/deepseek-v3.2",
    "messages": [
      {"role": "user", "content": "Explain this error in three sentences and propose the smallest patch."}
    ]
  }'

# Multi-file, terminal, or finished English: model = openai/gpt-5.4
# Video, audio, or Gemini multimodal: model = google/gemini-3.1-pro-preview
# Empty preview or dropped reasoning blocks: same prompt, fall back to openai/gpt-5.4

The route is deliberately conservative. It does not try to prove that one brand is stronger at everything. It avoids three expensive mistakes: summarizing a source that does not fit in 160K, forcing a preview model through a tool loop that drops reasoning blocks, and paying frontier output prices to generate a hundred first drafts. Prices were still moving in October 2026. DeepSeek’s output quote can differ by more than 2× across providers on the same page. Re-run the same 20 tasks once a month. That is safer than freezing a slug inside a quarterly plan.

FAQ