Migrating a production agent to GPT-5.6: 2.2× faster, 27% cheaper - the real post-mortem

In plain terms

The ploy.ai team migrated an agent from GPT-5.4 to GPT-5.6 in production and published the raw figures: 2.2× less latency, 27% less cost, +8 points in completion rate. The post-mortem is more useful than marketing benchmarks.

The fact

Published on July 12, 2026, on ploy.ai/blog, the writeup covers a real migration: a sales agent (research + qualification + email) handling ~40k requests/day. Progressive switch over 5 days, A/B on the harness, full telemetry.

Analysis

Three concrete lessons are worth remembering:

1. The latency gain comes mostly from the model, not the harness. The same agent trace runs 2.2× faster in median, 3.1× in P99. This is consistent with OpenAI's announcements on serving optimizations (speculative decoding + KV cache reuse).
2. The cost gain is partly illusory. The price per token has decreased, but the new version tends to generate longer outputs (~+18%). The net -27% comes first from the shortening of agent rounds (fewer iterations), not from the price per token alone.
3. The +8 points in completion rate comes from better tool-calling. On multi-step workflows, GPT-5.6 makes fewer tool mistakes and retries less often. This is a harness gain induced by the model, not a gain in raw intelligence.

Under the hood

The harness required three modifications for the switch: (1) adjust the system prompt (less explicit coaching as the model needs it less), (2) harden the tool call parser (new, stricter call format), (3) recalibrate retry thresholds (fewer failures → lower threshold).

The total compute budget only decreases if you fix the target latency - otherwise, the temptation is to use it to add steps.

Scenarios

• Wide adoption (60%): most production agents switch to 5.6 by Q4 2026, motivated by the latency/cost/tool-calling trio.
• Partial adoption (30%): those with a fine-tuned 5.4 model wait for the 5.6 fine-tuning pipeline.
• Wait-and-see (10%): those with strong vendor lock-in (Azure early tier, annual contracts) wait for renewal.

So what

For a team preparing the switch: the 5 days of A/B on the harness side are non-negotiable. The gains are real but depend on your harness - no shortcuts. For a manager: the real KPI is cost per successful workflow, not cost per token. The difference can be a factor of 2.

To watch

The post-mortems of teams that migrated to Claude Cowork or to Gemini 3 - objective baselines are still missing. The upcoming 5.6 fine-tuning pipelines (announced "summer" by OpenAI).