AI Development17 min readApril 2026

Qwen 3.6 Plus vs Claude Opus 4.6 vs GPT-5.4 Compared

Three frontier models with context windows of 1M tokens or more now compete for developer attention: Alibaba’s Qwen 3.6 Plus with thinking on by default and the lowest list price of the three, Anthropic’s Claude Opus 4.6 with adaptive thinking and 80.8% on SWE-bench Verified, and OpenAI’s GPT-5.4 with 75.0% on OSWorld-Verified, above the human score OpenAI cites. This comparison covers vendor-reported benchmarks, pricing and the trade-offs that decide which model fits your workload.

This dated chapter keeps its original context; our maintained frontier model guide covers current API prices, access limits, and workload selection.

Digital Applied Team

April 2, 2026• Updated October 21, 2026

17 min read

80.8%

Claude SWE-bench Verified

75.0%

GPT-5.4 OSWorld-Verified

1M+

All three context windows

$0.50

Qwen 3.6 Plus input / 1M

Key Takeaways

Coding Scores:: On SWE-bench Verified, Anthropic reports 80.8% for Claude Opus 4.6 and Qwen reports 78.8% for Qwen 3.6 Plus; OpenAI reports no Verified score for GPT-5.4 and gives 57.7% on SWE-bench Pro (Public)

Price Spread:: Qwen 3.6 Plus lists at $0.50/$3 per million tokens on Alibaba Cloud Model Studio (international, up to 256K input), GPT-5.4 at $2.50/$15 and Claude Opus 4.6 at $5/$25

Context Parity:: All three offer context windows of 1M tokens or more: GPT-5.4 at 1.05M, Claude Opus 4.6 at 1M (standard pricing since March 13, 2026) and Qwen 3.6 Plus at 1M

Computer Use:: OpenAI reports 75.0% on OSWorld-Verified for GPT-5.4, above the 72.4% human score it cites; on the same benchmark Anthropic reports 72.7% for Opus 4.6 and Qwen reports 62.5% for Qwen 3.6 Plus

Reasoning Controls:: Qwen 3.6 Plus thinks by default but can be switched off with enable_thinking; Claude Opus 4.6 uses adaptive thinking with four effort levels; GPT-5.4 sets reasoning.effort from none to xhigh

By early April 2026, three frontier models from three different labs offered context windows of 1M tokens or more. Qwen 3.6 Plus, Claude Opus 4.6 and GPT-5.4 each take a distinct approach to reasoning, coding and agentic work. This comparison will help you decide which model, or combination, fits your production needs.

Corrected October 21, 2026: we re-checked every figure against Alibaba Cloud, Qwen, Anthropic and OpenAI primaries. Qwen 3.6 Plus was a paid Model Studio API from its April 2 launch (the free tier was an OpenRouter preview), it thinks by default rather than always and accepts images and video; Claude Opus 4.6 scored 65.4% on Terminal-Bench 2.0, not 74.7%; and GPT-5.4 mini pricing and the OpenRouter usage ranks were wrong. We fixed those and removed unsourced speed multiples, estimated scores and a “first to beat humans” claim.

Comparison Date: April 2, 2026. Benchmark figures are each vendor’s own: Qwen’s April 2 launch post, Anthropic’s Claude Opus 4.6 announcement and system card, and OpenAI’s GPT-5.4 release post. Harnesses and settings differ, so read cross-vendor gaps as indicative. Qwen 3.6 Plus appeared as a free preview on OpenRouter on March 30 and launched on Alibaba Cloud Model Studio on April 2.

Update, October 2026: all three models have since been succeeded. Anthropic lists Claude Opus 4.6 as a legacy model (retirement not before February 5, 2027) behind Claude Opus 5.5, OpenAI released GPT-5.5 on April 23, 2026, and Alibaba Cloud Model Studio now lists Qwen 3.7 and Qwen 3.8 models. Prices and limits below are as of early April 2026.

Model Overview & Key Differences

Each model represents a different company’s vision for frontier AI. Alibaba pairs a low list price with image and video input. Anthropic emphasises safety, reliability and adjustable reasoning effort. OpenAI pushes on computer use and one model across ChatGPT, the API and Codex. All three return text only and none takes audio input; teams that need native audio understanding should weigh the leading omnimodal model contenders.

Qwen 3.6 Plus

Released: April 2, 2026 (OpenRouter preview from March 30)

Company: Alibaba (Qwen Team)

Context: 1M tokens / 65K output

SWE-bench Verified: 78.8% (Qwen-reported)

Pricing: $0.50/$3 per 1M tokens (international, up to 256K input)

Best for: Value, agentic coding, image and video input

Claude Opus 4.6

Released: February 5, 2026

Company: Anthropic

Context: 1M tokens / 128K output

SWE-bench Verified: 80.8%

Pricing: $5/$25 per 1M tokens

Best for: Complex coding, reliability, safety

GPT-5.4

Released: March 5, 2026

Company: OpenAI

Context: 1.05M (922K in / 128K out)

SWE-bench Pro (Public): 57.7%

Pricing: $2.50/$15 per 1M tokens

Best for: Computer use, unified capabilities

Architecture Approaches

Feature	Qwen 3.6 Plus	Claude Opus 4.6	GPT-5.4
Architecture	Not detailed by Alibaba (OpenRouter’s listing: hybrid linear attention + sparse MoE)	Not disclosed	Mainline reasoning model with GPT-5.3-Codex coding
Reasoning mode	Hybrid: thinking on by default, switchable (enable_thinking)	Adaptive thinking, effort low to max	reasoning.effort none (default) to xhigh
Input modality	Text + Image + Video	Text + Image	Text + Image
Max output	65K tokens	128K tokens	128K tokens
Model variants	Single (Plus)	Opus + Sonnet	gpt-5.4, Pro; mini and nano from March 17
Availability	Model Studio API + OpenRouter	API + claude.ai	API + ChatGPT + Codex

Benchmark Performance Comparison

Scores are as each vendor reported them at launch. No vendor ran all three models on the same harness, so each cell names its own conditions in the note below the table.

Benchmark	Qwen 3.6 Plus	Claude Opus 4.6	GPT-5.4
SWE-bench Verified	78.8%	80.8%	Not reported
SWE-bench Pro	56.6% (Qwen-corrected set)	Not reported	57.7% (Public set)
Terminal-Bench 2.0	61.6%	65.4%	75.1%
OSWorld-Verified	62.5%	72.7%	75.0% (human 72.4%)
GDPval	Not reported	GDPval-AA Elo 1606 (different metric)	83.0% wins or ties
OpenRouter usage rank (tokens)	#2 (free preview listing)	#7	Outside top 9

Qwen 3.6 Plus: Qwen’s April 2 launch post (SWE-bench on an internal bash and file-edit scaffold; SWE-bench Pro on a task set Qwen corrected; Terminal-Bench 2.0 on the Terminus-2 harness, average of five runs). Claude Opus 4.6: Anthropic’s announcement and system card (Terminal-Bench 2.0 on Terminus-2 at max effort; OSWorld-Verified, the name the system card uses, first-attempt success averaged over five runs; GDPval-AA is Artificial Analysis’s Elo version of GDPval). GPT-5.4: OpenAI’s release post (xhigh reasoning effort; Terminal-Bench harness not stated; GDPval share of comparisons won or tied against professionals). Usage rank: OpenRouter’s weekly token chart for the week beginning March 30, 2026, as archived on April 3, a measure of gateway traffic rather than capability. That chart counts Qwen 3.6 Plus’s free preview and free launch listings separately (second and third); the leaderboard list on the same archived page puts Claude Opus 4.6 sixth and GPT-5.4 outside the top 10.

Benchmark Context: SWE-bench Verified and SWE-bench Pro measure different difficulty levels, and no two of these models report both on the same terms. Claude Opus 4.6 has the highest reported Verified score; GPT-5.4 has the highest reported Pro score, on the public set, with Qwen 3.6 Plus 1.1 points behind on its own corrected set. On Terminal-Bench 2.0 (agentic terminal coding), as each vendor reports it, GPT-5.4’s 75.1% sits 9.7 points above Claude Opus 4.6’s 65.4% and 13.5 above Qwen 3.6 Plus’s 61.6%.

Speed & Throughput

Performance is not just about accuracy. Response speed matters for developer experience and production latency:

Speed Metric	Qwen 3.6 Plus	Claude Opus 4.6	GPT-5.4
Relative speed	No vendor figure	No vendor figure	Priority processing at 2x price
Streaming support	Yes	Yes	Yes
Batch API	Half price in Chinese Mainland deployment only	Yes (50% savings)	Yes (50% savings)
Prompt caching	Yes (context cache, discounted)	Yes (90% savings)	Yes ($0.25/1M cached)

Measure Speed on Your Own Prompts

None of the three vendors published a like-for-like latency or throughput figure at launch. Measure time to first token and output tokens per second on your own prompts, with Qwen 3.6 Plus thinking on and off, and with Claude Opus 4.6 and GPT-5.4 at the effort level you plan to run. OpenAI also sells faster GPT-5.4 responses through priority processing at twice the standard rate.

Pricing & Cost Analysis

Qwen 3.6 Plus entered at a fraction of its rivals’ per-token prices, and Anthropic dropped its long-context premium for Claude Opus 4.6 on March 13, 2026. For context on how these compare to open-source alternatives, see our companion guide.

Model	Input / 1M	Output / 1M	Cached Input	Typical Request*
Qwen 3.6 Plus (international, ≤256K input)	$0.50	$3.00	Discounted	$0.08
Qwen 3.6 Plus (256K–1M input)	$2.00	$6.00	Discounted	Variable
GPT-5.4	$2.50	$15.00	$0.25	$0.40
GPT-5.4 (>272K context)	$5.00 (2x)	$22.50 (1.5x)	$0.50	Variable
Claude Opus 4.6	$5.00	$25.00	$0.50	$0.75
Claude Opus 4.6 (batch)	$2.50	$12.50	$0.25	$0.375
Claude Sonnet 4.6	$3.00	$15.00	$0.30	$0.45
GPT-5.4 Mini	$0.75	$4.50	$0.075	$0.12

* Typical request: 100K input tokens, 10K output tokens. Grayed rows show cheaper alternatives within each family. Qwen 3.6 Plus prices are the international (Singapore) list prices on Alibaba Cloud Model Studio’s qwen3.6-plus page; Alibaba’s price list as archived in April 2026 gives the same tiers in yuan (¥3.7471 / ¥22.4826 and ¥14.9884 / ¥44.965). The Chinese Mainland (Beijing) deployment lists ¥2 / ¥12 per 1M tokens up to 256K input and ¥8 / ¥48 above it, and context-cache rates are discounted rather than published here. The GPT-5.4 surcharge applies to the full session once input exceeds 272K tokens. GPT-5.4 mini (March 17, 2026) has a 400K context window. Claude and OpenAI prices from Anthropic’s pricing docs and OpenAI’s model pages.

Key Pricing Insight: Since March 13, 2026, Anthropic has billed the full 1M window of Claude Opus 4.6 and Sonnet 4.6 at standard rates. OpenAI charges 2x input and 1.5x output for GPT-5.4 sessions whose input exceeds 272K tokens, which lifts it to $5 / $22.50: level with Opus 4.6 on input and still below it on output. Qwen 3.6 Plus also steps up above 256K input, to $2 / $6, and remains the cheapest of the three at every length.

Monthly Cost: 1,000 Requests (100K in / 10K out)

$80

Qwen 3.6 Plus (international)

$400

GPT-5.4

$750

Claude Opus 4.6

Claude batch API reduces cost to $375/mo. GPT-5.4 falls to $175/mo if the whole 100K-token prompt is a cache hit.

Context Windows & Reasoning

All three models offer context windows of 1M tokens or more, but context size is only part of the story: how each model prices and reasons within that context matters as much. For a broader view including open-source models with up to 10M context, see our comprehensive context window comparison.

Context Feature	Qwen 3.6 Plus	Claude Opus 4.6	GPT-5.4
Context window	1M tokens	1M tokens	1.05M tokens
Max output	65K tokens	128K tokens	128K tokens
Max input	991,808 tokens (983,616 with thinking)	1M window shared with output	About 922K (1.05M less 128K output)
Long-context surcharge	Higher tier above 256K ($2 / $6)	None (removed March 13)	2x input, 1.5x output above 272K

Reasoning Architecture Compared

How these models think is as important as how much they can read:

Qwen 3.6 Plus

Hybrid Thinking, On by Default

Alibaba Cloud documents Qwen 3.6 Plus as a hybrid thinking model: it reasons before answering unless you set enable_thinking to false. Its chain of thought is capped at 81,920 tokens, thinking_budget sets a lower cap, and the new preserve_thinking option keeps earlier turns’ reasoning for agent loops.

Trade-off: Thinking runs unless you turn it off, so simple requests cost more tokens and time by default.

Claude Opus 4.6

Adaptive Thinking

Claude decides when deeper reasoning helps. Developers steer how selective it is with four effort levels: low, medium, high (the default) and max, which Opus 4.6 introduced.

Trade-off: Effort guides rather than fixes the amount of thinking, so token use varies from request to request.

GPT-5.4

Effort Levels Plus Variants

In the API, reasoning.effort runs from none (the default) through low, medium and high to xhigh. GPT-5.4 Pro spends more compute on the hardest tasks, and GPT-5.4 mini and nano (March 17) trade capability for cost. GPT-5.4 Thinking is the ChatGPT name for the same model.

Trade-off: Picking both a model and an effort level per task adds routing decisions.

Coding Capabilities Deep Dive

Coding is the most contested benchmark category, with each vendor leading on a different metric. Here is how the vendor-reported coding results and tooling compare:

Coding Metric	Qwen 3.6 Plus	Claude Opus 4.6	GPT-5.4
SWE-bench Verified	78.8%	80.8%	Not reported
SWE-bench Pro	56.6% (Qwen-corrected set)	Not reported	57.7% (Public set)
Terminal-Bench 2.0	61.6%	65.4%	75.1%
Agentic coding	Launch focus (Qwen Code)	Claude Code	Codex
Front-end generation	QwenWebBench 1501.7 Elo (Qwen’s internal test; Opus 4.5 1517.9)	No vendor figure	No vendor figure
Code execution	Code interpreter tool	Code execution tool (sandbox)	Code interpreter tool

Scores and test conditions as in the benchmark table above. QwenWebBench is Qwen’s own front-end benchmark and Qwen did not test Opus 4.6 or GPT-5.4 on it. All three code tools run in a hosted sandbox: Anthropic’s code execution tool, OpenAI’s code interpreter in the Responses API, and Model Studio’s code interpreter, which Alibaba Cloud’s current documentation lists for the Qwen 3.6 Plus series.

Coding Strengths by Model

Qwen 3.6 Plus

Alibaba positioned the launch around agentic coding, from front-end pages to repository-level fixes, and reports 78.8% on SWE-bench Verified, close to Opus 4.5’s 80.9% in its own table, at the lowest price of the three.

Claude Opus 4.6

Highest reported SWE-bench Verified score of the three (80.8%), a measure of fixing real GitHub issues. Integrated with Claude Code for terminal-based development and MCP for tool use.

GPT-5.4

Highest reported scores on SWE-bench Pro (57.7%, public set) and Terminal-Bench 2.0 (75.1%), the harder multi-file and terminal tests. Brings GPT-5.3-Codex’s coding into the mainline model, with effort levels and mini and nano variants for cost tuning.

Developer Tool Integration

Terminal Agents

Claude Opus 4.6 runs Claude Code and GPT-5.4 runs Codex. Qwen documents Qwen 3.6 Plus setups for Qwen Code (1,000 free calls a day with Qwen sign-in), OpenClaw and Claude Code through its Anthropic-compatible endpoint.

IDE Integration

GitHub made Claude Opus 4.6 (February 5) and GPT-5.4 (March 5) generally available in Copilot on their launch days. Qwen also names Kilo Code, Cline and OpenCode as coding assistants that work with its API.

API Compatibility

All three support function calling and structured output. Model Studio serves Qwen 3.6 Plus through OpenAI-compatible Chat Completions and Responses APIs and an Anthropic-compatible API. GPT-5.4 uses OpenAI’s APIs and Claude the Anthropic Messages API.

Agentic & Tool Use Features

AI agents that can use tools, navigate computers, and execute multi-step workflows are the frontier of model capability in 2026. Each model approaches agency differently:

Agentic Feature	Qwen 3.6 Plus	Claude Opus 4.6	GPT-5.4
Function calling	Native	Native (plus MCP connector)	Native
Computer use	62.5% OSWorld-Verified	72.7% OSWorld-Verified	75.0% OSWorld-Verified
Tool protocol	OpenAI- and Anthropic-compatible	MCP (open standard)	OpenAI tools API, MCP servers
Code execution	Code interpreter tool	Code execution tool	Code interpreter tool
Agent platforms	Qwen Code, OpenClaw, Claude Code	Claude Code, GitHub Copilot	Codex, Copilot, ChatGPT
Structured output	Structured outputs	JSON schema outputs	JSON schema outputs

MCP Support: Anthropic introduced the Model Context Protocol (MCP), a standardized, open way to connect AI models to external tools, databases and APIs, and Claude Opus 4.6 supports it natively. MCP is no longer Anthropic-only: OpenAI’s Responses API accepts MCP servers for GPT-5.4, and Qwen reports MCPMark and MCP-Atlas scores for Qwen 3.6 Plus.

GPT-5.4: Highest Reported Computer-Use Score of the Three

OpenAI reports 75.0% on OSWorld-Verified for GPT-5.4, against 47.3% for GPT-5.2 and a human score of 72.4%. It is OpenAI’s first general-purpose model with built-in computer use in the API and Codex. Combined with hosted code execution and the tool search capability, GPT-5.4 is a strong fit for:

Automated testing and QA workflows requiring GUI interaction
Desktop automation for repetitive business processes
Cross-application workflows spanning multiple desktop tools

Which Model to Choose

The right frontier model depends on your priorities. Use this decision framework:

Choose Qwen 3.6 Plus When:

Budget is the primary concern: $0.50/$3 per million tokens up to 256K input is the lowest list price of the three
Front-end component generation and agentic coding are core workloads
Inputs include video, which Claude Opus 4.6 and GPT-5.4 do not accept
Your data can run in Alibaba Cloud’s Beijing, Singapore or US (Virginia) regions under Model Studio terms
A thinking on/off switch with an optional token budget suits you better than named effort levels

Choose Claude Opus 4.6 When:

The highest reported SWE-bench Verified score of the three (80.8%) matters most for your bug-fixing work
Long-context workloads benefit from 1M tokens at standard pricing (no surcharge)
Enterprise data policies, safety features, and established contractual agreements matter
Your agents are built around MCP and Claude Code
Adaptive thinking with a max effort level and 128K output tokens suit long-form code generation

Choose GPT-5.4 When:

Computer use and GUI automation are required: 75.0% on OSWorld-Verified is the highest reported score of the three
Five reasoning-effort settings plus Pro, mini and nano variants give fine cost/quality tuning per task
You want computer use, tool search and code interpreter as hosted tools in one API
Existing OpenAI ecosystem investment (Copilot, Azure OpenAI, ChatGPT Enterprise) should be leveraged
Multi-file and terminal coding (SWE-bench Pro 57.7%, Terminal-Bench 2.0 75.1%) and knowledge work (GDPval 83.0% wins or ties) are the primary workloads

Enterprise Deployment Considerations

For enterprise teams evaluating these models, several factors beyond raw performance matter:

Enterprise Factor	Qwen 3.6 Plus	Claude Opus 4.6	GPT-5.4
Data residency	Beijing, Singapore or US (Virginia) endpoints	US-only inference option (1.1x); cloud regions	Data residency endpoints (10% uplift)
Enterprise tier	Alibaba Cloud	Claude for Enterprise	ChatGPT Enterprise / Azure
SLA available	Per Alibaba Cloud agreement	Per enterprise agreement	Per enterprise agreement
SOC 2 / ISO 27001	Via Alibaba Cloud	Yes	Yes
Data training opt-out	OpenRouter’s free preview listing collected data; check Model Studio terms	Yes (API default)	Yes (API default)
Safety features	Not detailed at launch	RSP ASL-3 protections, cyber probes	Preparedness Framework (High cyber capability)

Preview Data Warning: OpenRouter’s free Qwen 3.6 Plus preview listing stated that prompts and completions are collected to improve the model. Do not send confidential, proprietary or client data through free endpoints. Use the paid Alibaba Cloud Model Studio API, in a region that fits your data rules, or Claude or GPT-5.4 for sensitive enterprise workloads.

Recommended Multi-Model Enterprise Strategy

Most enterprise teams benefit from a tiered approach that leverages each model's strengths:

Tier 1

High-Volume / Non-Sensitive

Qwen 3.6 Plus or GPT-5.4 Mini for draft generation, content summarization, and rapid prototyping.

Tier 2

Production Code / Critical Tasks

Claude Opus 4.6 or GPT-5.4 for code review, complex debugging and production changes, chosen by results on your own repositories.

Tier 3

Computer Use / Automation

GPT-5.4 for GUI automation, desktop testing, and cross-application workflow orchestration.

Ready to Integrate Frontier AI Models?

Whether you choose Qwen 3.6 Plus for value, Claude Opus 4.6 for reliability, or GPT-5.4 for computer use, our team can help you build the right AI infrastructure for your production needs.

Get Started Explore AI Services

Free consultation

Expert guidance

Tailored solutions

Sources

Benchmark scores are vendor-reported. Also named in the text without a link: OpenRouter’s model catalog (archived March 31 to April 10, 2026) and public rankings (archived April 3, 2026), GitHub’s changelog entries for Claude Opus 4.6 (February 5, 2026) and GPT-5.4 (March 5, 2026) in Copilot, and Artificial Analysis, which runs GDPval-AA. Llama 4 Scout’s 10M-token window is from Meta’s Llama 4 announcement (April 5, 2025).