Anthropic and OpenAI Launch Cheaper High-Performance Models Opus 5.5 and GPT-6 Sol

The frontier AI race shifted toward cost efficiency and practical engineering execution when Anthropic and OpenAI launched their newest models on the same day. On September 22, 2026, Anthropic unveiled Claude Opus 5.5, the first model in its Claude 5.5 family. As reported by TechCrunch, about 90 minutes later OpenAI released GPT-6 Sol, a mid-tier model in the GPT-6 family designed to bring reasoning to everyday developer workloads at lower cost.

While earlier release cycles emphasized peak benchmark records, this release cycle reflects a practical turning point for enterprise software. As Anthropic highlighted in its launch notes, cache reads make up the majority of agentic and coding workload costs, making sustained price-to-performance a central consideration for development teams running multi-step automation.

Claude Opus 5.5: About 40% Lower Typical Usage Cost

Anthropic positions Claude Opus 5.5 as its new primary workhorse for sprawling coding migrations, complex audits, and multi-step knowledge work. According to Anthropic's official announcement, Opus 5.5 performs at the level of Claude Fable 5.1 on most work while reducing typical token-billed usage costs by approximately 40% compared to Opus 5 at default settings. This is a vendor workload estimate, not a flat discount on every request.

Anthropic states that Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that reduced footprint. Because the model costs less per token and uses fewer tokens per task, the combination nets out to an estimated 40% drop in costs on typical workloads, while generating output more than 30% faster than Opus 5.

Metric	Claude Opus 5.5	Claude Opus 5	Generation Delta
Input Tokens (per 1M)	$4.00	$5.00	20% Cheaper
Output Tokens (per 1M)	$20.00	$25.00	20% Cheaper
Cached Input Reads (per 1M)	$0.20	$0.50	60% Cheaper
Cache Writes (per 1M)	$5.00 (5-min) / $8.00 (1-hr)	$6.25 (5-min) / $10.00 (1-hr)	20% Cheaper
Fast mode rates (research preview; first-party Claude API only)	$8.00 in / $40.00 out (up to 2.5x speed)	$10.00 in / $50.00 out (up to 2.5x speed)	20% Cheaper
Context Window	1,000,000 tokens	1,000,000 tokens	Identical
Max Output Tokens	128,000 tokens	128,000 tokens	Identical

Opus 5.5 defaults to "medium" effort adaptive thinking, which developers can adjust to trade off deliberation latency and cost. Thinking is always enabled on Opus 5.5; it cannot be switched off. Anthropic reported that Opus 5.5 was the strongest-performing model it had tested on its automated behavioral audit. In a separate containment evaluation measuring a model's propensity to cross containment boundaries, Anthropic found that Opus 5.5 attempted to circumvent boundaries around 85% less often than Opus 5 or Claude Mythos 5.1. Anthropic also cautions that evaluation awareness complicates predictions about real-world behavior. These results do not guarantee containment or establish a safety ranking against GPT-6 Sol.

Anthropic notes that developers can identify and fix bugs in their code as part of routine software development, but most cybersecurity tasks will be re-routed to Opus 4.8. For security practitioners, Anthropic plans to soon expand its Cyber Verification Program to include Opus 5.5, while vetted organizations whose biology work is impeded by standard safeguards can apply for access through the new Life Sciences Verification Program.

GPT-6 Sol: Lower-Cost Coding and Agentic Workflows

OpenAI launched GPT-6 Sol alongside GPT-6 Luna, positioning Sol for demanding coding and agentic work at lower token prices. Both models share Astra's improved communication style and build on its alignment work. Sol's standard input and output rates are 50% below GPT-5.6 Sol's current promotional rates. Luna's input rate is 50% lower, while its output rate falls from $1.20 to $0.50 per million tokens, a reduction of about 58%. The GPT-5.6 Sol reference says its promotional pricing remains available at least through November 21, 2026.

GPT-6 Sol is built for complex coding and agentic workflows, featuring a 1.05 million token context window. OpenAI's announcement says that on its internal factuality evaluation, which is based on de-identified real-world conversations where users flagged mistakes, GPT-6 Sol makes about half as many mistakes as its predecessor. OpenAI cautions that these error-inducing conversations are not representative of typical usage.

Feature / Spec	GPT-6 Sol	GPT-5.6 Sol (predecessor; promotional pricing)
Input Price (per 1M)	$2.00	$4.00
Output Price (per 1M)	$10.00	$20.00
Cached Input Read (per 1M)	$0.20	$0.40
Cache Writes (per 1M)	$2.50	$5.00
Reasoning Effort Range	none, low, medium (default), high, xhigh, max	none, low, medium (default), high, xhigh, max
Context Window Limit	1,050,000 tokens	1,050,000 tokens
Long-Context Surcharge Threshold	Above 272K input tokens (2x input/cache, 1.5x output)	Above 272K input tokens (2x input, 1.5x output)

OpenAI's reasoning guide documents how to append a configuration_update input item between responses, before the next user message, while keeping the request-level reasoning.effort unchanged. Sol supports effort settings from none to max. These updates preserve the earlier prefix for cache reuse and work in standard, single-agent mode. They cannot be combined with automatic compaction or automatic truncation; the standalone /responses/compact endpoint also rejects histories containing them.

The prompt-caching guide explains that changing tool definitions, ordering, schemas, or instructions can prevent reuse from the changed point onward. Earlier matching prefixes can still be reusable. Changing which tools are callable while keeping their definitions stable is a separate, cache-preserving option.

Vendor-Reported Benchmark Breakdown: Coding and Tool Execution

Neither vendor included the competitor's new release in its launch tables. The results below come from Anthropic and OpenAI, with secondary reporting identified where used. These are selected runs, not a controlled head-to-head comparison or a consistent best-score ranking. Effort settings, evaluation subsets, harnesses, and fallback behavior differ; unreported results do not establish a capability gap.

Benchmark Category	Benchmark Suite	Claude Opus 5.5	GPT-6 Sol	Baseline / Notes
Terminal Coding	Terminal-Bench 4.0	66.4% (xhigh)	Not Published	GPT-6 Astra: 57.9% (high), GPT-5.6 Sol: 37.3%
Software Engineering	FrontierCode v1.1 (Main)	54.4% (max)	49.3% (max, $2.14/task; Kingy AI chart reading)	Claude Fable 5.1: 50.3% (Anthropic launch figure; effort unspecified), GPT-6 Astra: 53.3% (max)
Editor Integration	CursorBench 4.0	57.8% (max)	Not Published	Claude Fable 5.1: 51.8% (max), GPT-5.6 Sol: 41.7%
Complex SWE	DeepSWE v1.1	Not Published	68.8% (max effort)	Claude Fable 5: 69.9% (xhigh), GPT-5.6 Sol: 72.7%
Business Automation	AutomationBench	40.0% (max, early access, no fallbacks)	33.2% (xhigh, $0.27/task)	GPT-6 Astra: 41.4% (max), Opus 5: 26.9%
Computer Use	OSWorld 2.0 (Partial)	81.8% (max)	64.4% (offline set, max, $3.25/task; Kingy AI chart reading)	Different subsets; OpenAI claims Astra remains the world's best model for computer use
Knowledge Work	GDPval-AA v2.1	1846 Elo (max)	Not Published	Opus 5: 1708 Elo, GPT-6 Astra: 1542 Elo
Scientific Reasoning	Terminal-Bench-Science 0.1	58.7% (max)	Not Published	GPT-6 Astra: 64.6%, Opus 5: 29.0%

For FrontierCode, Kingy AI's reading of OpenAI's chart data also reports Sol at 48.4% (xhigh) and Fable 5.1 at 48.7% (xhigh). OpenAI's prose describes Sol as matching Fable 5.1 at xhigh without giving a numerical score. The table uses Sol's stronger max-effort run; the Fable 5.1 baseline is Anthropic's reported figure, which does not state an effort level.

Anthropic's general evaluation setup enabled production safeguards, with some tasks routed to older Claude models. Its AutomationBench launch result is an exception: Zapier ran that early-access evaluation without fallback models and counted safeguard interventions as failures. As checked on September 26, Zapier's public leaderboard lists a separate Opus 5.5 run at 42.47% with default fallbacks and max effort, costing $1.44 per task. That should not be conflated with the 40.0% launch run.

Anthropic cautions that small benchmark differences are becoming less useful predictors of practical performance. In its internal HAProxy rewrite from C into Rust, Opus 5.5 finished in 9.5 hours, passed nearly all regression tests, and cost 51% less than Fable 5.1.

Kingy AI's analysis reports Sol's best DeepSWE result as 68.8% at max effort and $2.74 per task, below GPT-5.6 Sol's 72.7% and Opus 5's 73.7%. Kingy interprets this as a trade-off: Sol reduces spending substantially while giving up some accuracy at the highest-scoring setting. On Zapier's AutomationBench suite, which tests end-to-end workflows using 47 tools, GPT-6 Sol achieved a 33.2% completion rate at xhigh effort and $0.27 per task.

In DataCamp's hands-on test building a single-file Tetris clone with reversing gravity, both models passed the core logic, with DataCamp noting: "GPT-6 Sol wins this test, but narrowly, and not on the hard part", citing Sol's visual polish and responsiveness. This single task does not establish a general coding advantage.

The Hidden Mechanics of Pricing and Workload Costs

While GPT-6 Sol has a 50% lower list price on headline token rates ($2/$10 vs $4/$20 per million tokens), DataCamp's workload modeling highlights that real-world cost differences depend heavily on context length and cache utilization.

The examples use standard API rates and equal token counts for both models. Different tokenizers, reasoning-token usage, retries, cache misses, and tool fees can change an actual bill. Batch, Fast mode, and regional pricing require separate calculations.

1. Long-Context Surcharges

OpenAI applies a long-context multiplier once a request exceeds 272,000 input tokens. Above this threshold, input and cache rates double, while output rates increase by 1.5x for the entirety of that request ($4.00 input, $15.00 output, $0.40 cache read).

Anthropic imposes no long-context surcharges, maintaining $4 input and $20 output rates per million tokens across Opus 5.5's full 1M context window. In DataCamp's modeled retrieval example, 10M uncached input tokens and 1M output tokens are spread across requests that each exceed 272K input tokens. Sol costs $55, versus $60 for Opus: an 8.3% saving for this token mix. Above the threshold, Sol's cached reads cost $0.40 per million tokens versus Opus's $0.20. With enough cached input, Opus can cost less despite its higher output rate.

2. Prompt Cache Read Parity Below the Surcharge Threshold

At standard rates, with Sol requests at or below 272K input tokens, both models charge $0.20 per million cached input tokens. DataCamp's cache-heavy example writes a 100K prefix once, then reuses it in 1,000 requests, each with 5K uncached fresh input and 1K output tokens. Sol totals $40.25, versus $60.50 for Opus.

This assumes one successful prefix write at the listed rate (the 5-minute rate for Claude), cache hits throughout, and no additional cache writes. Sol's total is $0.25 for the write + $20 for reads + $10 for fresh input + $10 for output. Opus's total is $0.50 + $20 + $20 + $20. Sol is 33.5% cheaper in this example; Opus costs about 50% more. The saving changes with the token mix.

Ecosystem Availability and API Integration

Both models launched through first-party APIs and enterprise cloud providers. Availability below reflects published information checked September 26, 2026; access depends on product, plan, region, rollout, and workspace settings.

Integration Surface	Claude Opus 5.5	GPT-6 Sol
API Model Identifier	claude-opus-5-5	gpt-6-sol
First-Party Interfaces	Claude Pro, Max, Team, Enterprise	ChatGPT Work and Codex on Plus, Pro, Business, Enterprise, and Edu; rollout and workspace controls apply
Cloud Platforms	Amazon Bedrock, Google Cloud, Microsoft Foundry	Microsoft Foundry, Amazon Bedrock
Developer Tools	Claude Code, Cursor, GitHub Copilot	Codex CLI, GitHub Copilot
Third-Party Gateways	OpenRouter, Vercel AI Gateway	OpenRouter, Vercel AI Gateway

Anthropic's model documentation lists Opus 5.5 on Amazon Bedrock, Google Cloud, and Microsoft Foundry, and Cursor lists it among its supported models. OpenAI's official changelog describes a gradual rollout of Sol in Work and Codex, with Enterprise administrators required to enable new models. Work is a product surface, not a subscription tier; Sol is not yet available in Chat. OpenAI also documents Sol access through Amazon Bedrock.

GitHub's launch notice lists Sol for Copilot Pro+, Max, Business, and Enterprise, with a gradual rollout and administrator controls. DataCamp reports no Sol availability in Vertex AI or Cursor at publication; Cursor's model list still does not list Sol as of this check.

Strategic Recommendation: Choosing Between Opus 5.5 and GPT-6 Sol

Use these results to shortlist models, then compare task completion, latency, and total cost on your own workload:

• Consider Claude Opus 5.5 if:

  • Your workflows involve long coding tasks, repository migrations, or architectural refactoring, and you want to test whether its reported strengths carry over to your codebase.
  • Your requests frequently exceed 272,000 input tokens, especially with heavy cache reuse. Compare its flat rates against Sol's long-context rates; context size alone does not determine the cheaper model.
  • Your infrastructure requires availability through Amazon Bedrock, Google Cloud, and Microsoft Foundry, or your developers build heavily inside Cursor.
  • You value Anthropic's reported containment improvements and can validate the model and its safeguards against your own deployment requirements.

• Consider GPT-6 Sol if:

  • You run high-volume, cost-sensitive automation pipelines, customer-facing workflows, or batch triage systems where requests stay at or below 272K input tokens and lower token rates translate into lower cost per completed task.
  • Your development team already standardizes on OpenAI Codex and ChatGPT Work environments.
  • You want dynamic reasoning control that can drop down to none or low effort for simple steps and scale up to xhigh or max for difficult steps, within the API compatibility limits described above. Opus 5.5 has no equivalent to Sol's none effort setting: its adaptive thinking is always enabled.
  • You need lower-priced business-tool automation and can check completion rates and response times against your requirements.

The release of Claude Opus 5.5 and GPT-6 Sol expands the range of viable architectures for software teams. Lower token rates can make longer agent runs more affordable, but the useful comparison is cost per successfully completed task, including retries, reasoning, caching, and tool usage.

---

Sources & References

Pricing, documentation, and availability checked September 26, 2026. Benchmark dates and conditions are identified above.

• Anthropic Announcement: Introducing Claude Opus 5.5 (Accessed September 2026)
• Anthropic Documentation: Fast mode (research preview) and Pricing (Accessed September 2026)
• OpenAI Announcement: Introducing GPT-6 Sol and Luna (Published September 22, 2026)
• OpenAI Documentation: GPT-6 Sol Model, GPT-5.6 Sol Model, Pricing, Prompt caching, Reasoning models, and ChatGPT & Codex changelog (Accessed September 2026)
• Zapier: AutomationBench leaderboard and methodology (Checked September 26, 2026)
• TechCrunch: OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes (Published September 22, 2026)
• DataCamp: GPT-6 Sol vs Claude Opus 5.5: Benchmarks, Pricing, and Which to Use (Published September 23, 2026)
• Kingy AI: GPT-6 Sol and GPT-6 Luna: Specs, Benchmarks, Pricing and How They Compare to Claude Opus 5.5, Fable 5.1 and Gemini (Published September 22, 2026)