By the PromptUnit team·Published September 30, 2026·Last updated September 30, 2026·Facts checked October 16, 2026·9 min read

OpenAI Doubled the Frontier Price. Claude Opus 4.7 Took the Coding Lead a Week Earlier.

GPT-5.5 launched at $5/$30 per million tokens, double GPT-5.4. Claude Opus 4.7 sits at $5/$25 with the SWE-Bench lead. Here's what changed for routing.

llm cost optimizationcross-provider llm routinganthropic api costopenai api costmodel routing llm

OpenAI launched GPT-5.5 on April 23, 2026 at $5 per million input tokens and $30 per million output tokens. That is exactly double the rate of GPT-5.4. Seven days earlier, on April 16, Anthropic shipped Claude Opus 4.7 at $5 per million input and $25 per million output, and it scored 64.3% on SWE-Bench Pro and 87.6% on SWE-Bench Verified, beating every public OpenAI model on coding benchmarks.

For the first time in two years, the cheapest path to frontier-grade coding output runs through Anthropic, not OpenAI. If your routing logic still defaults coding traffic to OpenAI by reflex, the math against your bill has already changed.

This is not a tribal argument. It is a pricing and benchmark walk-through that any engineering team running real LLM volume should run against their own traffic.

What launched, when, and at what price

Claude Opus 4.7 shipped on April 16, 2026. Pricing held flat from Opus 4.5 at $5 per million input tokens and $25 per million output tokens. The 1M context window remained, vision support extended to 2,576-pixel images, and the headline numbers across coding, agentic reasoning, and knowledge benchmarks all moved up.

GPT-5.5 shipped on April 23, 2026. ChatGPT Plus, Pro, Business, and Enterprise got it first; the public API is described as "coming very soon" pending additional safety guardrails for serving at scale. Pricing landed at $5 per million input tokens and $30 per million output tokens, with a 1M context window. That is exactly 2x the price of GPT-5.4 on both axes. There is also a Pro variant at $30 input and $180 output, plus Batch and Flex tiers at half the standard rate and Priority at 2.5x.

For anyone tracking frontier API costs over time, this is the largest single-step price increase on a flagship OpenAI model in three years.

The benchmarks

On SWE-Bench Pro, the standard for autonomous software engineering tasks, Claude Opus 4.7 scored 64.3% against GPT-5.5's 58.6%. That is a 5.7-point Anthropic lead.

On SWE-Bench Verified, Opus 4.7 scored 87.6%, up from 80.8% on Opus 4.5. This is the strongest publicly reported number for autonomous coding agents on this benchmark.

On MCP Atlas, which tests Model Context Protocol fluency for multi-tool orchestration, Opus 4.7 scored 77.3%. That number is best-in-class across all generally available models.

GPQA Diamond, the graduate-level reasoning benchmark, gave Opus 4.7 a 94.2%. CharXiv visual navigation hit 79.5% at full resolution, up from 57.7% on Opus 4.5. Finance Agent landed at 64.4%, also state-of-the-art.

GPT-5.5 wins where it always has: agentic browsing, knowledge work, tasks that mix text and images, and tasks where OpenAI's tooling stack already integrates deeply. On Terminal-Bench, GDPval, OSWorld, and τ2-bench, GPT-5.5 still leads. The split, in other words, is real and stable: Anthropic owns coding and tool orchestration; OpenAI owns the broader agentic and knowledge surface.

OpenAI's cost defense

OpenAI's public defense of the GPT-5.5 price hike is that the model uses about 40% fewer output tokens to complete the same Codex task as GPT-5.4. If that holds in production, the effective per-task cost increase is closer to 20% than 100%.

That is a meaningful claim and worth taking seriously. But it is also the kind of claim that is true on average and false in 30% of edge cases, and it does not change the per-token sticker price on the bill. If your traffic is dominated by long-output generation tasks where token efficiency drops, you will pay closer to the full 2x. If your traffic is dominated by terse tool-call sequences, you will pay closer to the 20%. The honest answer is that you have to measure your own workload before assuming you got the average.

Update: GPT-5.5 is now available on the public API at the $5/$30 rate described above, so the apples-to-apples comparison against Opus 4.7 is now measurable in production rather than theoretical. GPT-5.4 at $2.50/$15 remains the cheaper OpenAI option for workloads that don't need the newer model.

The 2x price jump is easier to defend for some workloads than for others. GPT-5.5 takes text and images and answers in text, so it suits tasks that mix the two. Whether that is worth routing to depends heavily on the workload: see the GPT-5.5 multimodal cost breakdown.

Let each request use the cheapest model that's just as good

PromptUnit checks each request your app sends. When a cheaper model is just as good for that request, it uses it; otherwise your model answers. Free until it has saved you money.

Try it free →

What changed about cross-provider routing

Cross-provider routing has always been technically possible. The question has been whether it was worth the operational overhead. Different APIs, different rate limit behaviors, different failure modes, different prompt formats. We covered why the calculus has tilted toward yes for most production workloads in our piece on cross-provider LLM routing.

These two launches sharpened the calculus on coding workloads specifically.

Before April 16, the default routing logic for a coding agent was: send everything to GPT-5.4, fall back to Claude Sonnet 4.6 for cost, escalate to Opus 4.5 for hard cases. The default model and the cheap model both lived inside OpenAI's stack. Cross-provider routing was an option, not a requirement.

After April 23, the math looks different. The cheapest path to frontier-grade coding output is Opus 4.7 at $5/$25. The next-cheapest path is GPT-5.4 at $2.50/$15, but it is now two generations behind on coding benchmarks. The new flagship from OpenAI, GPT-5.5, costs more output-side than Opus 4.7 ($30 vs $25) and loses the coding benchmarks. So if your workload is coding-heavy and quality matters, you have one provider winning on benchmarks at lower output cost than the OpenAI flagship. Open-weight models such as DeepSeek V4 Pro list lower prices than either, but their coding scores are vendor-reported; see our DeepSeek V4 Pro page.

This is the exact scenario cross-provider routing was built for, and it just became the dominant case for any coding-heavy production workload.

What to do with this

If you are running a coding agent, a code review pipeline, an IDE assistant, or any workflow where SWE-Bench-style task quality is the bottleneck, now is the time to add Anthropic as a primary route, not a fallback. The path looks like this:

Route to Opus 4.7: hard coding tasks, multi-file refactors, autonomous agent loops with five or more steps, tool-orchestration heavy work, and anything requiring deep reasoning over long context.

Route to GPT-5.4 mini: bounded coding tasks, classification, intent detection, structured extraction, and all the other categories we covered in our analysis of when the cheaper model wins after the GPT-5.4 mini release.

Run your own token-efficiency tests before routing meaningful volume to GPT-5.5. The 40% fewer output tokens claim is plausible, but production workloads vary, and the 2x sticker price means you cannot afford to take the average on faith.

Keep GPT-5.4 in the mix for non-coding agentic work, knowledge synthesis, and tool-use tasks where OpenAI still leads on Terminal-Bench, OSWorld, and τ2-bench. That decision has not changed, and the routing logic should reflect it.

The deeper pattern

There is a longer arc here that engineering teams should pay attention to. For two years, the default frontier model assumption was OpenAI, and routing decisions inherited that default. The result was the default-flagship trap we wrote about in the hidden cost of defaulting to GPT-4o in production: paying flagship rates for traffic that did not need flagship capability.

April 2026 is the first month where that default has visibly broken. The OpenAI flagship doubled in price. The Anthropic flagship took the coding crown without raising prices. The mini-tier OpenAI model approached flagship benchmarks at one-third the cost. Each of these shifts independently would justify revisiting routing logic. Together, they make any team that is still on a single-provider, single-model setup measurably overpaying.

The fix is not to switch providers. The fix is to route per request to the cheapest model that meets the quality bar for that specific task. The model lineup has gotten too good and too price-stratified for static defaults to keep working.

How a routing proxy handles this

PromptUnit is a proxy you point your app's AI API base URL at. It checks each request and uses a cheaper model only when that model is just as good for that request; otherwise the request goes to the model you asked for. The dashboard shows each request with the model you asked for, the model that answered, the cost and the saving. It is free until it has saved you $5 (individual) or $100 (business, work email), then it takes 20% of the savings.

If your team is running a coding-heavy LLM workload and has not benchmarked Opus 4.7 against your current default, that is the highest-leverage routing change available right now. Start at promptunit.ai.

For the complete cross-provider routing strategy, see Cross-Provider LLM Routing. For the Claude Haiku vs Sonnet routing guide to optimize Anthropic spend, see Claude Haiku vs Sonnet: Routing Guide. For the full model routing framework, see LLM Model Routing: The Complete Guide. For Anthropic's closed frontier tier and what Claude Mythos pricing means for routing decisions, see Claude Mythos: What It Costs Now, and Why the Launch Price Didn't Stick. For the GPT-5.4 mini vs flagship routing math, see GPT-5.4 vs GPT-5.4 Mini: The Routing Decision OpenAI Made For You.

Try It Free

See exactly where your AI budget is going. PromptUnit shows every request with the model that answered, what it cost and what it saved, starting with your first one.

Start free, no credit card needed. Or talk to us if you want a walkthrough.