The MeshTok catalog lists more than a dozen models with a context window of one million tokens or more. Almost all of them trade something away to get there — no vision, no video, no audio, a small max output, or a missing flagship tag. Gemini 3.1 Pro is the exception: a 1,048,576-token context window, a 65,536-token max output, native image+video+audio input, reasoning, tools, json, and the flagship label, all in one model id. The question this article answers is whether that combination is worth the $2 / $12 per-million-token price.
This is a spec-and-pricing analysis, not a synthetic benchmark. We did not run latency tests, did not measure needle-in-a-haystack recall at each context position, and we are not going to publish any number we cannot back up with the live catalog. Every figure below is arithmetic you can reproduce with a calculator from src/data/models.ts. The goal is to help you decide whether Gemini 3.1 Pro’s million-token window earns a slot in your routing layer — based on the data that actually matters for budgeting: context window, max output, multimodal coverage, and per-million-token price.
One honest note before the numbers: the model id in the catalog is google/gemini-3.1-pro-preview. The -preview suffix is in the source data, and we will not hide it. Treat the spec sheet as accurate, but verify against https://meshtok.com/v1/models before locking a production routing decision.
Gemini 3.1 Pro: the spec sheet
Available today through the MeshTok unified API at https://meshtok.com/v1 under the model id google/gemini-3.1-pro-preview.
| Field | Value |
|---|---|
| Vendor | Google (US) |
| Context window | 1,048,576 tokens (2²⁰, slightly above 1M) |
| Max output | 65,536 tokens |
| Input price (per 1M) | $2.00 |
| Output price (per 1M) | $12.00 |
| Capabilities | image, video, audio, reasoning, tools, json, flagship |
| Released | 2026-06 |
| Knowledge cutoff | 2026-05 |
Three things stand out at this tier:
- The largest combined context+output window of any multimodal model in the catalog. 1,048,576 input + 65,536 output means a single call can ingest roughly a million tokens of mixed text, images, audio, and video, and emit a 65K-token response. No other model in
models.tssimultaneously matches all three: 1M+ context, 65K+ max output, and image+video+audio input. - Full native multimodal, including video and audio. Most 1M-context competitors stop at
image(DeepSeek V4 Pro, Qwen3.7 Max, Doubao Seed 2.1 Pro) or drop vision entirely (GLM-5.2). Gemini 3.1 Pro is the only 1M-context flagship in the catalog that accepts video and audio natively alongside text and images. - It is labeled
flagshipandreasoning. Google applies the same capability tags it reserves for its frontier tier. The cheaper Gemini Flash variants carryfastinstead ofreasoningandflagship— that is the structural divide inside the family.
The million-token question: what does 1M context actually cost?
This is the part most marketing pages skip. A million-token context window is a budget, not a freebie. Every token you put in is billed at the input rate. The honest way to compare long-context models is to compute the cost of actually using the window.
Cost of one fully-loaded call (1M input + a short summary)
A common long-context workload: send ~1M tokens of documents and ask for a short summary or a few action items. We assume 1,048,576 input tokens and 1,000 output tokens.
| Model | Input cost | Output cost | Total per call |
|---|---|---|---|
| Gemini 3.1 Pro | $2.097 | $0.012 | $2.109 |
| Gemini 2.5 Pro | $1.311 | $0.010 | $1.321 |
| Gemini 3 Flash | $0.524 | $0.003 | $0.527 |
| DeepSeek V4 Pro | $0.449 | $0.001 | $0.450 |
| GLM-5.2 | $1.143 | $0.004 | $1.147 |
| Claude Sonnet 5 | $3.000 | $0.015 | $3.015 |
A single fully-loaded Gemini 3.1 Pro call costs about $2.11. Run that a thousand times a day and you are spending $2,109/day — roughly $63K/month — on input tokens alone. That is not a reason to avoid the model; it is a reason to be deliberate about when you actually need the full window.
Cost of a fully-loaded long-output call (1M input + full max output)
Where Gemini 3.1 Pro’s pricing becomes structurally different is when you also use the large max output. Here we assume the context is filled to the brim and the model is allowed to emit its full max output.
| Model | 1M input | Max output used | Output cost | Total per call |
|---|---|---|---|---|
| Gemini 3.1 Pro | $2.097 | 65,536 | $0.786 | $2.883 |
| GLM-5.2 | $1.143 | 128,000 | $0.512 | $1.655 |
| DeepSeek V4 Pro | $0.449 | 32,768 | $0.028 | $0.477 |
| Claude Sonnet 5 | $3.000 | 16,384 | $0.246 | $3.246 |
GLM-5.2 is cheaper per fully-loaded call because its max output is 128K but priced at $4/1M, not $12/1M. The deciding factor between Gemini 3.1 Pro and GLM-5.2 is not price — it is whether you need video and audio input, which GLM-5.2 does not accept at all.
Cost of 1 million short chat calls (1K input + 500 output each)
For a sanity check at the other end of the spectrum: one million short calls, each consuming 1,000 input tokens and producing 500 output tokens. That is 1 billion input tokens and 500 million output tokens in aggregate.
| Model | Input cost (1B in) | Output cost (500M out) | Total for 1M calls |
|---|---|---|---|
| Gemini 3.1 Pro | $2,000 | $6,000 | $8,000 |
| Gemini 2.5 Pro | $1,250 | $5,000 | $6,250 |
| Gemini 3 Flash | $500 | $1,500 | $2,000 |
| DeepSeek V4 Pro | $429 | $429 | $857 |
| GLM-5.2 | $1,143 | $2,000 | $3,143 |
| Claude Sonnet 5 | $3,000 | $7,500 | $10,500 |
On short chat traffic, Gemini 3.1 Pro is about 9.3× more expensive than DeepSeek V4 Pro and 2.5× more expensive than GLM-5.2. The lesson is simple: the million-token window is a premium feature, and you should not pay for it on traffic that never uses it.
Inside the Gemini family: which one to pick?
The MeshTok catalog carries seven Gemini models with a 1M-class context window. Picking the right one is mostly a tradeoff between reasoning depth, multimodal coverage, and price.
| Model | Context | Max out | Input $/M | Output $/M | Video | Audio | Reasoning | Flagship |
|---|---|---|---|---|---|---|---|---|
| Gemini 3.1 Pro | 1,048,576 | 65,536 | $2.00 | $12.00 | ✅ | ✅ | ✅ | ✅ |
| Gemini 2.5 Pro | 1,048,576 | — | $1.25 | $10.00 | ✅ | ✅ | ✅ | ✅ |
| Gemini 3.5 Flash | 1,048,576 | 65,536 | $1.50 | $9.00 | ✅ | ✅ | — | — |
| Gemini 3 Flash | 1,048,576 | 65,536 | $0.50 | $3.00 | ✅ | ✅ | — | — |
| Gemini 3.1 Flash Lite | 1,048,576 | — | $0.25 | $1.50 | ✅ | ✅ | — | — |
| Gemini 2.5 Flash | 1,048,576 | — | $0.30 | $2.50 | ✅ | ✅ | — | — |
| Gemini 2.5 Flash Lite | 1,000,000 | — | $0.10 | $0.40 | ✅ | ✅ | — | — |
Reading the matrix:
- Gemini 3.1 Pro vs Gemini 2.5 Pro — 2.5 Pro is the older flagship and is cheaper on both axes ($1.25/$10 vs $2/$12). The case for paying more for 3.1 Pro: a defined 65,536 max output (2.5 Pro’s entry has no explicit
maxOutputin the catalog), a more recent knowledge cutoff (2026-05), and the newer 3.1 generation. The case for staying on 2.5 Pro: if you do not need those, it is strictly cheaper. - Gemini 3.1 Pro vs Gemini 3 Flash — Flash is one-quarter the price and shares the same 1M context, 65K max output, and full multimodal input. What you give up is
reasoningandflagship. For high-volume multimodal classification, transcription, or routing where step-by-step thinking is not needed, Flash is the right pick. For anything that benefits from reasoning chains — multi-step analysis, long-document Q&A with synthesis — 3.1 Pro is the upgrade. - Gemini 3.1 Pro vs Gemini 3.1 Flash Lite / 2.5 Flash Lite — the Lite variants are the budget end. They keep the 1M context and multimodal input but drop reasoning and the explicit max-output cap. Use them when you need to ingest a lot of tokens cheaply and the response quality bar is moderate.
Gemini 3.1 Pro vs other 1M-context flagships
How does Gemini 3.1 Pro stack up against the better-known 1M-context flagships from other labs?
| Model | Context | Max out | Input $/M | Output $/M | Video | Audio | Released |
|---|---|---|---|---|---|---|---|
| Gemini 3.1 Pro | 1,048,576 | 65,536 | $2.00 | $12.00 | ✅ | ✅ | 2026-06 |
| DeepSeek V4 Pro | 1,000,000 | 32,768 | $0.4286 | $0.8571 | — | — | 2026-04 |
| GLM-5.2 | 1,000,000 | 128,000 | $1.1429 | $4.0000 | — | — | 2026-07 |
| Qwen3.7 Max | 1,000,000 | 8,192 | $0.8571 | $2.5714 | — | — | 2026-06 |
| Claude Sonnet 5 | 1,000,000 | 16,384 | $3.0000 | $15.0000 | — | — | 2026-05 |
| Claude Opus 4.8 | 1,000,000 | 32,000 | $5.0000 | $25.0000 | — | — | 2026-06 |
Two structural observations:
- Gemini 3.1 Pro is the only model in this table with native video and audio input. Every other 1M-context flagship accepts images at most. If your workload is video understanding, meeting transcription, or any pipeline that ingests audio/video alongside text, Gemini 3.1 Pro is not the cheapest option — it is the only option in the 1M-context tier.
- GLM-5.2 has the largest max output (128K) but no multimodal at all. If your workload needs to generate very long single-turn text responses from text-only input, GLM-5.2 is both cheaper and has a longer output window. The moment you need an image, video, or audio, GLM-5.2 is off the table and Gemini 3.1 Pro re-enters the conversation.
The long-context vs RAG tradeoff (the honest answer)
The question “is the million-token window worth it?” almost always reduces to a cost question, not a capability question. A 1M context lets you skip a retrieval pipeline and just dump everything into the prompt. That is convenient, but at $2/1M input it is also expensive on every single call — and you pay it again on the next call, because the context window does not cache for free.
A simple rule of thumb, derived purely from the pricing above:
- If the same large corpus is queried many times, retrieval is cheaper. Send 5K retrieved tokens at $2/1M = $0.01 per call instead of 1M tokens at $2/1M = $2.00 per call. You break even after roughly 200 retrieved-only calls for every one full-context call.
- If the corpus is queried once or twice and then discarded (one-shot summarization of a fresh transcript, a single deep-read of a contract, ad-hoc exploration of a codebase), long context is the right tool. The retrieval system would cost more to build than the tokens it saves.
- If you need audio or video understanding over a long timeline (a 2-hour meeting recording, a long video), Gemini 3.1 Pro is the only model in the catalog that combines that input modality with a 1M window. There is no cheaper substitute inside the 1M tier — the tradeoff is between Gemini 3.1 Pro and a shorter-context model plus chunking.
When to pick Gemini 3.1 Pro (and when not to)
There is no “best model” — only the right model for a given workload and budget. Use this matrix as a starting point and validate with your own eval.
| Workload | Recommended | Why |
|---|---|---|
| Long video or audio understanding (1M+ tokens of media) | Gemini 3.1 Pro | Only 1M-context model in the catalog with native video and audio input. No cheaper substitute in this tier. |
| Mixed multimodal long-context analysis (text + images + audio) | Gemini 3.1 Pro | Only flagship combining 1M context, 65K max output, and full multimodal input. |
| Long-context text-only with very long output (>65K, up to 128K) | GLM-5.2 | GLM-5.2’s 128K max output is unique. Cheaper on fully-loaded long-output calls. No vision. |
| High-volume 1M-context multimodal on a budget | Gemini 3 Flash | One-quarter the price of 3.1 Pro, same 1M context, same multimodal input. Drops reasoning. |
| Cheapest possible 1M context (text or image) | DeepSeek V4 Pro | $0.43/1M input — 4.7× cheaper than Gemini 3.1 Pro on input. No video/audio. |
| Top-tier coding quality, cost-secondary | Claude Sonnet 5 | Anthropic’s positioning and community evals put Sonnet 5 at the top for code. Gemini 3.1 Pro if multimodal long context matters more. |
| Long-context analytical reasoning where multimodal is not needed | GLM-5.2 or DeepSeek V4 Pro | Both cheaper for text-only long-context work. Pick GLM-5.2 for long output, V4 Pro for lowest cost. |
A reasonable default routing policy: route any call that contains video or audio to Gemini 3.1 Pro; route long-text-only refactors with very long output to GLM-5.2; route cheap high-volume text+image traffic to DeepSeek V4 Pro; reserve Claude Sonnet 5 for high-stakes coding. Gemini 3.1 Pro is the multimodal long-context specialist, not the default for everything.
How to call Gemini 3.1 Pro (real, runnable code)
MeshTok exposes a single OpenAI-compatible endpoint at https://meshtok.com/v1. Switching between Gemini 3.1 Pro and any other model in this article is a one-line change to the model field — same SDK, same API key, same request format. The example below shows the long-context multimodal pattern: a long text prompt plus an image URL, sent as a single call.
# pip install openai
from openai import OpenAI
client = OpenAI(
base_url="https://meshtok.com/v1",
api_key="sk-your-meshtok-key",
)
def ask(model_id: str, prompt: str, image_url: str | None = None) -> str:
"""Same call, different model and modality — switch with one string."""
content = [{"type": "text", "text": prompt}]
if image_url:
content.append({"type": "image_url", "image_url": {"url": image_url}})
resp = client.chat.completions.create(
model=model_id,
messages=[{"role": "user", "content": content}],
)
return resp.choices[0].message.content
# Real model IDs from the MeshTok catalog
multimodal_long = ask("google/gemini-3.1-pro-preview",
"Summarize this 800K-token transcript and extract action items.",
image_url="https://example.com/chart.png")
text_only_long = ask("bigmodel/glm-5.2",
"Rewrite this 50K-token file with the new API.")
cheap_long = ask("deepseek/deepseek-v4-pro",
"Summarize this 200K-token document.")
coding_flagship = ask("anthropic/claude-sonnet-5",
"Refactor this Python function and add tests.")Streaming, tool calling, JSON mode, and image input all work the same way — change the model string and add the relevant parameters. For streaming responses, pass stream=True; for tool calling, pass the tools argument; for structured output, use response_format={"type": "json_object"}. See the MeshTok streaming docs and the tool-calling docs for full examples in Python, JavaScript, and cURL.
FAQ
Is Gemini 3.1 Pro really the only 1M-context model with video and audio input?
As of 2026-07-21, yes. Every other model in src/data/models.ts with a 1M+ context window — DeepSeek V4 Pro, DeepSeek V4 Flash, GLM-5.2, Qwen3.7 Max/Plus, Qwen3.5 Plus/Flash, Qwen3.6 Plus, Claude Sonnet 5, Claude Opus 4.6/4.7/4.8, MiniMax M3, Longcat 2.0, Mimo V2 Pro/V2.5 — accepts either text only or text plus images. None carries the video or audio capability tag. The only models with native video/audio input are the Gemini family itself.
Why no latency or recall benchmarks in this article? Because we did not run any, and we are not going to publish numbers we cannot back up with raw logs. Long-context recall (the “needle in a haystack” question) depends on prompt shape, position, language, and the model’s attention behavior at each depth — a single recall percentage without the test harness, the prompts, and the raw outputs is marketing, not data. If you need recall numbers for your workload, run them yourself on MeshTok with your own corpus. The pricing and capability data above is the part that is stable and verifiable.
Is the -preview suffix in the model id a problem?
It is a signal, not a blocker. Google ships 3.1 Pro as a preview model in the catalog (google/gemini-3.1-pro-preview). That usually means the spec sheet is accurate but the model may change behavior without a version bump. For production routing where stability matters more than the latest capability, Gemini 2.5 Pro is the non-preview flagship alternative at a lower price.
How does Gemini 3.1 Pro compare to Gemini 2.5 Pro?
2.5 Pro is the older Google flagship. It is cheaper ($1.25/$10 vs $2/$12), carries the same flagship + reasoning + full multimodal tags, and has the same 1,048,576 context window. The case for paying more for 3.1 Pro: a defined 65,536 max output, a newer knowledge cutoff (2026-05), and the newer 3.1 generation. The case for staying on 2.5 Pro: it is strictly cheaper if you do not need those three things.
Does Gemini 3.1 Pro support tool calling and structured outputs?
Yes. The capabilities field lists tools and json, which means native function calling and JSON-mode structured output are both supported through the MeshTok OpenAI-compatible endpoint. The request format is identical to OpenAI’s — only the model field changes.
Is the 1M context window real or marketing?
It is the declared context window in src/data/models.ts, mirrored from Google’s own model card. Real usable context depends on your prompt shape, retrieval strategy, and how the model handles long-context attention — all of which you should test on your own data. The catalog figure is the upper bound the model accepts, not a quality guarantee at every position.
Why is Gemini 3.1 Pro so much more expensive than DeepSeek V4 Pro? Different cost structures and different capabilities. DeepSeek V4 Pro is aggressively priced for text+image volume and does not accept video or audio. Gemini 3.1 Pro is priced as a full multimodal flagship with native video and audio support — that is a different input surface, not a like-for-like text comparison. MeshTok does not add any markup; you see exactly what each lab publishes.
Are these prices stable?
Model vendors update list prices periodically. The figures in this article reflect the MeshTok catalog on 2026-07-21. The updatedDate at the top of the page tells you when the article was last reviewed; the live source of truth is always https://meshtok.com/models.
Last reviewed 2026-07-21 by the MeshTok Team. All prices and capabilities are taken from the live MeshTok model catalog (src/data/models.ts) on the review date. No latency, token-count, recall, or quality measurements were taken for this article — figures shown are arithmetic projections from published list prices, not empirical benchmarks.