GPT-5.6 Agent Migration: 27% Cheaper
Ploy says its production agent moved to GPT-5.6 Sol and became 2.2x faster and 27% cheaper than Claude Opus 4.8.
By AI Pricing Guru Editorial Team
AI Pricing Guru articles are maintained by the editorial workflow behind the site: daily pricing snapshots, provider source checks, and review passes for model launches, subscription limits, and billing changes.
Ploy says it moved its production website-building agent from Claude Opus 4.8 to GPT-5.6 Sol after internal evals showed finished builds running 2.2x faster and 27% cheaper.
That is not an official OpenAI price cut. GPT-5.6 Sol still sits at the premium OpenAI rate in our live data: $5 input, $0.50 cached input, and $30 output per 1M tokens. The pricing story is better than a sticker-price change, though: a production agent team says the same workflow became cheaper because GPT-5.6 used fewer tokens, finished faster, and only beat Opus after the team rebuilt its tool schemas and prompt-cache strategy.
For buyers comparing model economics, this is the useful lesson: frontier model ROI is no longer just input price times output price. It is model behavior, cache design, tool-call shape, retries, and cost per accepted task.
For current rates, keep the OpenAI pricing page and Anthropic pricing page open, then run your own workload through the AI token calculator. For broader coding-agent context, see our best AI for coding pricing guide and OpenAI vs Anthropic API pricing comparison.
What Changed
Ploy’s July 9 migration writeup says its agent now runs on GPT-5.6 Sol by default. The agent is a real production workflow, not a single chat completion: it plans websites, reads codebases, writes components, generates imagery, screenshots its own output, and decides when a build is done.
In Ploy’s redesign eval sample, the reported mean per completed build was:
| Metric | Claude Opus 4.8 | GPT-5.6 Sol |
|---|---|---|
| Cost | $3.06 | $2.22 |
| Wall-clock time | 8m 00s | 3m 42s |
| Input tokens | 2.60M | 1.70M |
| Output tokens | 33.0K | 17.1K |
| Visual score | 0.936 | 0.970 |
That works out to roughly 27% lower cost and a little over 2.2x faster wall-clock completion in the reported sample.
The important qualifier is that the numbers came after migration work. Ploy says early eval failures were not just model-quality problems. Some were harness assumptions, tool-schema differences, file-read bugs, and prompt-cache configuration issues that penalized GPT-5.6 until the team fixed them.
Pricing Context
The headline model price still looks expensive:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-5.6 Sol | $5.00 / 1M | $0.50 / 1M | $30.00 / 1M |
| GPT-5.6 Terra | $2.50 / 1M | $0.25 / 1M | $15.00 / 1M |
| GPT-5.6 Luna | $1.00 / 1M | $0.10 / 1M | $6.00 / 1M |
| Claude Opus 4.8 | $5.00 / 1M | $0.50 / 1M | $25.00 / 1M |
On sticker price alone, GPT-5.6 Sol does not beat Claude Opus 4.8. Input and cached-input pricing match Opus 4.8 in our dataset, while GPT-5.6 Sol output is higher.
The reported savings came from workload behavior:
| Driver | Pricing impact |
|---|---|
| Lower input tokens | Fewer paid prompt tokens per finished build |
| Lower output tokens | Less expensive generated code and reasoning text |
| Faster completion | Better product latency and less wall-clock agent time |
| Fewer failed reads after schema fixes | Less wasted retry traffic |
| Better prompt-cache design | More repeated context billed at cached-input rates |
That is why the post matters for AI pricing. It is a production example where the higher-output-price model can still win if it needs materially fewer tokens and fewer tool loops to finish the job.
The Cache Lesson
Ploy’s most useful cost finding is about prompt caching.
The team says its agent has a static prefix of roughly 29,000 tokens across tool schemas and system instructions. Under Claude, explicit cache controls let that static prefix cache broadly across the organization. Ploy reported Claude cache hit rates around 92% to 96%.
The GPT-5.6 migration initially looked about 50% more expensive, according to the post, because the cache was wrong. Ploy says GPT-5.6 required explicit prompt_cache_breakpoint markers and a prompt_cache_key; per-conversation keys missed shared context, while one global key ran into cache-node throughput limits. The fix was a workspace-scoped key with breakpointed prompt layers.
After that change, Ploy says first-call cache hits rose from roughly 0% to 83.7%, total uncached input tokens dropped 28%, and GPT-5.6 landed below Opus on the eval cost.
For buyers, the practical read is simple: do not compare GPT-5.6 and Claude costs until both sides have production-grade caching. A cold cache can make the wrong model look expensive.
The Tool-Call Lesson
Ploy also found a provider-specific tool-call behavior that affected cost and quality.
Its code tool had 25 top-level parameters. Claude usually sent only the few parameters it used. GPT-5.6 sent all 25 parameters every time, including plausible-looking values for unused optional fields. Ploy says that caused 52% to 64% of GPT-5.6 file reads to come back empty in one failure mode, even though the tool returned success.
The fix was to transform optional properties into required-but-nullable fields for OpenAI-family models, then strip nulls before validation. After the change, Ploy says empty file reads dropped to 0% and the agent needed roughly 30% fewer tool calls.
This is exactly where hidden AI costs come from. A model can be capable, but if it interacts badly with your tool schema, you pay for bad reads, retries, wasted reasoning, and extra recovery turns.
What This Means For Buyers
If you run coding agents, website agents, data agents, browser agents, or long-running tool-using workflows, GPT-5.6 Sol deserves a real eval. But the eval has to measure finished work, not just token price.
Track:
| Metric | Why it matters |
|---|---|
| Cost per accepted task | Best single pricing metric for agents |
| Input, cached input, and output tokens | Separates raw price from cache quality |
| Tool-call count | Finds hidden loops and schema friction |
| Retry rate | Shows whether a cheaper model is actually cheap |
| Wall-clock completion time | Matters for user-facing agent products |
| Human cleanup time | Captures the cost outside the API bill |
The Ploy example also argues against lazy migrations. If you simply swap model IDs, GPT-5.6 may look worse than it is. If you tune schemas, caching, reasoning replay, and eval harness assumptions, it may become cheaper even at premium rates.
Practical Advice
Use GPT-5.6 Sol as an escalation candidate for difficult agent work, not as a blind default everywhere. Its $5/$30 rate still needs justification.
Benchmark GPT-5.6 Terra and Luna alongside Sol. If the lower tiers preserve enough quality for parts of the workflow, routing can cut costs further.
Audit your cache before drawing conclusions. For long system prompts, tool schemas, repository context, brand guidelines, or workspace memory, cached-input economics can decide the winner.
Normalize tool schemas by provider. Optional fields, nullable fields, enum handling, array limits, and strict-mode behavior can change both quality and spend.
Finally, compare against Claude Opus 4.8 and Claude Sonnet 5 on your own tasks. Ploy’s website-agent workload is valuable evidence, but it is still one workload. The right answer for support automation, legal review, code repair, or RAG may differ.
Bottom Line
Ploy’s GPT-5.6 migration is a pricing event because it shows how production agent costs can fall without an official price cut.
GPT-5.6 Sol is not cheap on the rate card. But in Ploy’s reported agent eval, it used fewer input tokens, generated about half the output tokens, finished more than twice as fast, and came in 27% cheaper than Claude Opus 4.8 after cache and tool-schema fixes.
For AI buyers, the lesson is to price agents by finished work. Token rates matter, but the winning model is the one that produces accepted output with the least total waste.
Sources: Ploy’s GPT-5.6 migration writeup, OpenAI GPT-5.6 preview, and AI Pricing Guru’s live pricing dataset.