[!NOTE] Executive Summary (TL;DR)
- 2026 Frontier Landscape: Comprehensive comparison across frontier closed models (GPT-5.4, Claude 4.6, Gemini 3) and open-weights leaders (DeepSeek-V4, Kimi K3) across pricing and benchmarks.
- Rigorous Fact Checks: Confirms official Claude Opus 4.6 pricing at $5/$25 per million tokens, and verifies GPT-5.4 modalities (text and vision inputs without direct fine-tuning support).
- Specialized Deep Dive: For national open-source model ecosystems, consult the 2026 Domestic LLM Comprehensive Benchmark Report.
Scope and evidence
This comparison preserves the GPT-5.4, Claude Opus / Sonnet 4.6, and Gemini 3.1 Pro group. It is an engineering reference, not a claim that these are today's newest models. Specifications come from provider documentation; unsupported score tables are not treated as controlled measurements. See the Chinese open-weight comparison.
1. Capabilities and API boundaries
| Model | Input and output | Context | Engineering checks |
|---|---|---|---|
| GPT-5.4 | Text / image input, text output; no direct audio input | 1.05M total window, up to 128K output; shared input / output budget | Reasoning effort, Responses tools, computer use |
| Claude Opus / Sonnet 4.6 | Text / image input; document processing supports PDF | Consult each model's 1M-context conditions | Adaptive thinking, tool parameters, available model IDs |
| Gemini 3.1 Pro Preview | Multimodal understanding; check exact supported input formats | Limits depend on the Preview ID | Media accounting, tools, thinking configuration |
Sources: GPT-5.4, Opus 4.6, Sonnet 4.6, Gemini 3.1 Pro Preview.
ChatGPT's Thinking label is not a separate gpt-5.4-thinking API ID. Claude Code Agent Teams is a harness capability rather than automatic multi-agent orchestration in the base API. Distinguish API thinking parameters from separate consumer product modes.
2. Traceable token prices
Standard examples below are USD per million tokens. Account separately for long-context rates, cache writes, batch processing, regions, and tools.
| Model | Input | Output | Cached input |
|---|---|---|---|
| GPT-5.4 | 2.50 | 15.00 | 0.25 |
| Claude Opus 4.6 | 5.00 | 25.00 | 0.50 cache read |
| Claude Sonnet 4.6 | 3.00 | 15.00 | 0.30 cache read |
| Gemini 3.1 Pro Preview | Check input-length tier | Check input-length tier | Check cache and storage rates |
Sources: OpenAI model pricing, Claude model documentation, Gemini pricing. GPT-5.4 applies long-context multipliers above 272K input tokens, so the standard rate cannot be applied to every million-token request.
3. Fair reasoning, coding, and agent comparisons
SWE-bench Verified, HumanEval, OSWorld, and mathematics tests measure different capabilities. Do not place Opus results in a Sonnet column or describe provider-specific harness results as one uniform reproduction.
Pin the snapshot, reasoning budget, context, tool permissions, agent harness, timeout, and retries. Save patches, execution logs, and independent acceptance results. Compare quality, latency, and useful cost; see LLM evaluation.
| Workload | Acceptance evidence |
|---|---|
| Repository changes | Hidden tests, patch scope, repeated-run stability |
| Document extraction | Numeric accuracy, citations, retrieval, refusals |
| Computer operation | Reset environment, success rate, unauthorized actions, duration |
| Video / audio understanding | Supported modality, temporal grounding, quality, cost |
A large window does not establish complete repository understanding. A reported hallucination reduction on one evaluation is not a fixed improvement for all workloads.
4. API examples
These examples require provider SDKs, credentials, and account access. Site verification does not call external models.
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-5.4",
input="Explain the basics of quantum computing",
reasoning={"effort": "high"},
max_output_tokens=4096,
)
print(response.output_text)
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-sonnet-4-6",
max_tokens=4096,
messages=[{"role": "user", "content": "Explain quantum computing"}],
)
print(response.content)
from google import genai
client = genai.Client()
response = client.models.generate_content(
model="gemini-3.1-pro-preview",
contents="Explain quantum computing",
)
print(response.text)
Check IDs against your account's model list. Provider routing also needs consistent errors, deadlines, tool schemas, and usage accounting.
5. Self-hosting arithmetic and KV budgets
At an assumed $30/hour server cost and blended $5 per million API tokens, a simplified break-even point is 30 / 5 × 1e6 / 3600 = 1,667 tokens/s. This is not measured capacity or equivalent task quality. Input / output mix, utilization, expansion, and staffing change the result. See deployment cost estimation.
For ordinary GQA with 80 layers, 8 KV heads, head_dim=128, and FP16, cache bytes per token are 2×80×8×128×2=327,680. One million tokens require about 327.68GB or 305.18GiB. This is a defined example, not a rule for all 70B or hybrid models. PagedAttention improves allocation and reuse, not the size of an uncompressed KV element.
Frequently Asked Questions (FAQ)
Q1: Which model is strongest at coding?
Compare matching repository tasks, tools, and budgets. Provider benchmarks screen candidates; independent acceptance determines fit.
Q2: Does a million-token window guarantee accurate retrieval?
No. Capacity, retrieval quality, citation reliability, and cost need separate measurements.
Q3: How should cost and reliability be balanced?
Set minimum quality and SLA first, then compare useful cost. The token calculator helps separate input, output, and cache charges.
Originally published at LLM Systems Architecture on blog.llmgo.top.
Top comments (1)
A useful way to choose among these models is to look beyond benchmark scores and consider deployment fit. Test them on your own workload using the same prompts, tool permissions, retry policy, and latency budget. Then track task success, cost per successful run, and the severity of failures. This matters especially for agentic coding, where a model with a slightly lower headline score may still be the better choice if it uses tools more predictably and costs less to retry.
For the final user-facing copy, Wordfully's AI writing checker can help refine the vocabulary and wording. Keep that writing-quality pass separate from the model evaluation so it does not blur the benchmark results.