Claude Opus 5 vs GPT-5.6: 1M Context & Launched Benchmarks

Claude Opus 5 vs GPT-5.6: 1M Context, Honeycomb Codename, and the SWE Pro Claim

Claude Opus 5 is Anthropic's newly launched flagship — released July 24, 2026, using the pre-launch codename Honeycomb, with a 1M-token context window and configurable effort settings up to "max" — while GPT-5.6 Sol remains OpenAI's shipped incumbent. Both are now available with published pricing and hands-on benchmarks. Opus 5 leads on ARC-AGI-3, Frontier-Bench, OSWorld 2.0, and Artificial Analysis's Intelligence Index; GPT-5.6 Sol still leads DeepSWE. The early "beats Sol on SWE Pro" leak was directionally right on agentic and reasoning benchmarks.

Key Takeaways

• Claude Opus 5 launched July 24, 2026. Model ID claude-opus-5, priced at $5/$25 per million input/output tokens — the same as Opus 4.8 and half of Fable 5.
• GPT-5.6 remains shipped and measurable. Sol, Terra, and Luna variants are live with published Ultra Mode and multi-agent APIs — see our GPT-5.6 deep dive.
• Context window: 1M tokens confirmed for Opus 5 per Artificial Analysis and Anthropic's docs. GPT-5.6 context depends on tier.
• Head-to-head benchmarks: Opus 5 leads ARC-AGI-3 (30.2% vs 7.8%), OSWorld 2.0 (70.6% vs 62.6%), and Frontier-Bench (43.3% vs Opus 4.8's 21.1%). GPT-5.6 Sol still leads DeepSWE at 72.7%.
• Intelligence Index: Opus 5 (max) scores 61, narrowly ahead of Fable 5 (60) and GPT-5.6 Sol (59), at 26% lower cost per task than Fable 5.
• Launch context: Anthropic shipped Opus 5 the day after the widely-expected July 23 window slipped.

Claude Opus 5 vs GPT-5.6 at a Glance

Dimension	Claude Opus 5 (launched Jul 24, 2026)	GPT-5.6 (shipped)
Status	Live on Claude API, Claude Code, Pro/Max/Team/Enterprise, Bedrock, Vertex, Foundry	Generally available across Sol, Terra, Luna
Context window	1M tokens (confirmed)	Set by OpenAI per tier — see OpenAI docs
Reasoning modes	Effort settings from low → max; Fast Mode research preview (2.5x speed at 2x cost)	Ultra Mode (confirmed)
Coding benchmarks	Frontier-Bench 43.3% (vs 21.1% for Opus 4.8); DeepSWE trails Sol	DeepSWE 72.7% (leader); published SWE-bench numbers from OpenAI
Agentic benchmarks	ARC-AGI-3 30.2%, OSWorld 2.0 70.6%, BrowseComp 90.8%, AutomationBench 26.0%	ARC-AGI-3 7.8%, OSWorld 2.0 62.6%
Pricing (per M tokens)	$5 input / $25 output (same as Opus 4.8, half of Fable 5)	Set by OpenAI
API model ID	claude-opus-5	Live in OpenAI API
Agent focus	Long-running agents, per-turn effort controls, safety fallback to Opus 4.8	Multi-agent APIs

Capabilities

Claude Opus 5's shipped capability profile centers on long-running agents, advanced coding, and professional knowledge work. Anthropic positions it as a "daily driver" model that comes close to Fable 5's frontier intelligence at half the price. The pre-launch Honeycomb sighting in Cursor — the "Anthropic research model with per-turn controls and safety fallbacks" — landed as claude-opus-5 with the same architecture: effort settings from low through max, and a safety fallback that resolves to Opus 4.8 when needed.

Source: @LuminaXspace

GPT-5.6 ships three named variants — Sol, Terra, and Luna — with an Ultra Mode reasoning tier and multi-agent APIs, per our earlier GPT-5.6 deep dive. OpenAI has also folded Codex into the ChatGPT product, sharpening the coding-agent story further; see Codex absorbs ChatGPT for the mechanics.

Early hands-on impressions from Anthropic's Alex Albert are that Opus 5 is more token-efficient than Fable 5 for many coding tasks. Ivan Fioravanti reported it as "another level" compared to Opus 4.8 or Sonnet 5. Max Weinbach called Opus 5 and GPT-5.6 Sol "basically the same model" in day-to-day use — the benchmark gap is wider than the felt gap.

Benchmarks

Anthropic's launch benchmarks give the first head-to-head numbers between Opus 5 and GPT-5.6 Sol. The spread is real:

• ARC-AGI-3 (novel problem solving): Opus 5 scored 30.2%, roughly 4x GPT-5.6 Sol's 7.8%. Analysts noted Opus 5 used explicit algebraic reasoning to convert layouts into notation, a capability not previously seen from frontier models.
• Frontier-Bench (terminal coding): 43.3% for Opus 5, more than double Opus 4.8's 21.1%.
• OSWorld 2.0 (computer use): 70.6% vs GPT-5.6 Sol's 62.6%, beating Fable 5 at roughly one-third of Fable's cost.
• BrowseComp (agentic search): 90.8%.
• AutomationBench (business workflows): 26.0%, roughly 1.5x the next-best model at comparable cost.
• Humanity's Last Exam: 64.7%, narrowly edging Claude Mythos 5 at 64.5%.
• DeepSWE: GPT-5.6 Sol still leads at 72.7%.

Artificial Analysis independently placed Opus 5 (max) at 61 on the Intelligence Index — the highest score they've recorded, narrowly ahead of Fable 5 (60), GPT-5.6 Sol (59), Kimi K3 (57), and Opus 4.8 (56). Opus 5 also set new highs on GDPval-AA v2 (1861 Elo, +114 over Fable 5) and AA-Briefcase (1720 Elo, +146 over Fable 5), Artificial Analysis's agentic-knowledge-work benchmarks. See the Artificial Analysis comparison for the full breakdown.

The pre-launch ChrissGPT claim that Opus 5 would "beat 5.6 Sol & Terra on SWE Pro" was directionally right on agentic and reasoning benchmarks but wrong on pure software engineering: Sol still leads DeepSWE. Fable 5 also narrowly edges Opus 5 on FrontierCode and Humanity's Last Exam without tools, and leads on legal and health.

Caveats: Bindu Reddy noted Opus 5 sits just below Fable on LiveBench, arguing the model is bench-maxed on first-turn public benchmarks. Peter Gostev observed Opus 5's refusal rate on API-level checks has risen to around 10% (previous Opus models were near 0%; Fable was at 35%), suggesting some Fable-style safety classifiers have been ported over.

Pricing & Context Limits

Pricing is the sharpest surprise. Anthropic priced Opus 5 at $5 per million input tokens and $25 per million output tokens — identical to Opus 4.8, and exactly half of Fable 5 ($10 / $50). The Bindu Reddy worst-case rumor of 2x Opus 4.8 pricing did not materialize.

Anthropic also offers up to 90% savings with prompt caching and 50% with batch processing. A Fast Mode research preview provides up to 2.5x faster generation at double the standard cost ($10 in / $50 out), Claude API only for now.

One practical caveat from community testing: Opus 5 tends to spend roughly 2x the output tokens of Opus 4.8 at matched effort settings, so effective cost per task can rise even at identical per-token pricing. Users are advised to drop an effort notch when migrating.

On context: Anthropic and Artificial Analysis confirm a 1M-token context window, matching Fable 5 and Sonnet 5. GPT-5.6's context limit varies by tier and is set by OpenAI.

Availability

GPT-5.6: live in the OpenAI API and ChatGPT product. Available today.

Claude Opus 5: live as of July 24, 2026. Access surfaces:

• Claude API as claude-opus-5 (see Anthropic's launch post).
• Claude Code via /model claude-opus-5. Existing workloads can migrate via /claude-api migrate in Claude Code.
• Claude apps — new default model on Claude Max, strongest model on Claude Pro, available on Team and Enterprise.
• Cloud partners: Amazon Bedrock, Google Cloud (Vertex AI), and Microsoft Foundry.

The pre-launch signal chain — the Honeycomb EAP appearance in Cursor on July 9, the claude-opus-5 string on Vertex AI on July 14, the missed July 20-21 and July 23 windows — resolved when Andrew Curran and others reported Opus 5 was live in the Claude apps on the afternoon of July 24. Full background on the codename and pre-launch leak chain is in our Claude Opus 5 explainer.

One rough edge at launch: Anthropic's migration guide notes that web fetch is currently unavailable on Opus 5, so some API workflows can do less than they could on Opus 4.8.

Which One Should You Use?

Choose Claude Opus 5 if:

• You need frontier-level reasoning on novel problems — the ARC-AGI-3 gap over Sol (30.2% vs 7.8%) is real.
• Your workload is long-horizon agentic coding, computer use, or business workflows — Opus 5 leads OSWorld, BrowseComp, and AutomationBench.
• You need Fable-adjacent quality without paying Fable prices; $5/$25 undercuts Fable's $10/$50 by half.
• You want a 1M-token context with configurable effort tiers.

Choose GPT-5.6 if:

• You're doing straight software-engineering benchmarks — Sol still leads DeepSWE (72.7%).
• You want multi-agent APIs and Ultra Mode with a published feature surface.
• You're already invested in the OpenAI or Codex tooling stack, or want the ChatGPT product surface.

Choose neither yet — stay on Claude Opus 4.8 or GPT-5.5 — if: you're wary of Opus 5's roughly-10% API-level refusal rate, its missing web fetch tool on day one, or its higher token spend at matched effort. Fable 5.1 is also rumored within weeks, which may change the top-tier picture again.

Frequently Asked Questions

Is Claude Opus 5 better than GPT-5.6? On Anthropic's launch benchmarks, Opus 5 leads GPT-5.6 Sol on ARC-AGI-3 (30.2% vs 7.8%), OSWorld 2.0 (70.6% vs 62.6%), and Frontier-Bench. GPT-5.6 Sol still leads DeepSWE (72.7%). Artificial Analysis puts Opus 5 (max) at 61 on its Intelligence Index vs 59 for GPT-5.6 Sol (max), narrowly the most intelligent model tested. Real-world impressions are closer than the benchmark gap suggests.

Is Claude Opus 5 cheaper than GPT-5.6? Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens — identical to Opus 4.8 and half of Fable 5. Compare against OpenAI's official pricing page for GPT-5.6. Anthropic also offers a Fast Mode research preview at $10 in / $50 out for up to 2.5x faster generation.

What context window does Claude Opus 5 have vs GPT? Claude Opus 5 ships with a 1 million token context window, confirmed on Artificial Analysis and matching Claude Fable 5 and Claude Sonnet 5. GPT-5.6 context limits are set by OpenAI and vary by tier.

When was Claude Opus 5 released? Anthropic launched Claude Opus 5 on July 24, 2026. It is available today on the Claude API as claude-opus-5, in Claude Code, on Claude Pro / Max / Team / Enterprise, and via Amazon Bedrock, Google Cloud, and Microsoft Foundry.

What was Claude Honeycomb and was it Opus 5? Honeycomb was the pre-launch codename for the model that appeared briefly in Cursor's model picker on July 9, 2026 with a 1M-token context, an "xhigh" reasoning effort mode, and safety fallbacks to Claude Opus 4.8. It shipped as Claude Opus 5 on July 24.

Should I choose Claude Opus 5 or GPT-5.6 for coding? Opus 5 leads on Frontier-Bench terminal coding (43.3%, more than double Opus 4.8) and matches or beats Fable 5 on most coding tasks at half the price. GPT-5.6 Sol still leads DeepSWE. For agentic long-horizon coding and computer-use workloads, Opus 5 is the strongest single pick today; for straight DeepSWE-style tasks, Sol remains competitive.

Does Claude Opus 5 support agentic workflows better than GPT? Opus 5 is Anthropic's most agentic Opus to date, with per-turn effort controls and safety fallback routing. It scored 90.8% on BrowseComp for agentic search and 26.0% on AutomationBench for business workflows, leading GPT-5.6 Sol on both. GPT-5.6 Sol ships with multi-agent APIs and Ultra Mode and remains the incumbent for teams already invested in the OpenAI stack.

What to Watch Next

Three signals will settle the next round: (1) an updated Fable 5.1, rumored within weeks, which could reclaim the top of Anthropic's own lineup; (2) independent LiveBench and long-running agentic evaluations that stress-test the benchmark-vs-real-world gap flagged by Bindu Reddy and Peter Gostev; (3) whether Anthropic restores web fetch on Opus 5 and tunes the API-level refusal rate closer to Opus 4.8's near-zero baseline.

Building similar reasoning and coding agents? On kie.ai you can try GPT-5.6, Claude Opus 5, and Claude Sonnet 5.