OpenAI Just Killed Claude's Cost Advantage. GPT-5.6 Terra Matches Sonnet Pricing in September—And Wins on Integration.
On July 9, OpenAI shipped GPT-5.6: three model tiers (Sol, Terra, Luna) with pricing that mirrors Claude's, and integration that doesn't exist in Claude's API. Claude Sonnet 5 is $2 input / $10 output per million tokens through August 31. On September 1, that becomes $3 / $15. OpenAI's GPT-5.6 Terra is $2.50 / $15 permanently. By next month, Claude is 50% more expensive on input tokens, and GPT-5.6 ships with ChatGPT Work: an agent framework that ties reasoning directly to document generation, spreadsheets, browser automation, and scheduled tasks.
If your client pitch has been "Claude is cheaper and better at long-context," that pitch just evaporated.
The model tiers: one size does not fit all anymore
OpenAI shipped three GPT-5.6 models, each with explicit performance targets:
Sol ($5 input / $30 output): frontier capability. SOTA on coding benchmarks (88.8% on Terminal-Bench 2.1, a point above Claude Opus), cybersecurity reasoning, scientific problem-solving. Intended for tasks where quality is the constraint, not cost.
Terra ($2.50 / $15): the "sweet spot." Balanced intelligence and cost. Intended for knowledge work, writing, reasoning tasks where you're not pushing the frontier but you need more than basic completion. Benchmarks show Terra performs within 5-10% of Sol on most tasks.
Luna ($1 / $6): high-volume efficiency. Fast, cheap, designed for retrieval-augmented generation, summarization, and tasks where you can tolerate occasional hallucinations because the human is reading the output.
Claude hasn't shipped a tiered model family. You pick Opus or Sonnet or Haiku. Sonnet is the "all-purpose" model: decent at everything, best at nothing. GPT-5.6 Terra explicitly is the all-purpose model, with pricing and performance locked in. Starting September 1, that model costs less than Claude's all-purpose model.
ChatGPT Work: the integration moat
Here's the harder part: OpenAI didn't just ship models. They shipped ChatGPT Work: an agent that stays with a project for hours, creates documents, slides, spreadsheets, browses the web, and accesses a desktop app. The agent can spawn sub-tasks, schedule follow-up work, and interact with your productivity stack without bouncing to an API.
Claude ships APIs. You write code to call them. That's powerful. That's also a barrier for clients who don't want to hire engineers. ChatGPT Work removes the barrier. Your client's procurement team doesn't ask "is Claude cheaper?" anymore. They ask "does the agent solve my problem?" If the answer is "yes, and it integrates with our Slack," they buy ChatGPT Work.
Claude has Claude Cowork: a similar agent framework. But Cowork is positioned as a solo operator tool. ChatGPT Work is positioned as enterprise automation. One is "build something," the other is "get work done."
What this means for AI consulting
If you've been positioning yourself as "I integrate Claude and save you money," you have about six weeks before that positioning breaks.
The clients you were winning weren't actually choosing Claude because of the model. They were choosing Claude because:
- It was cheaper than GPT-4 or GPT-4 Turbo
- It had longer context windows than competitors
- Your integration was novel to them
Starting September, cheaper is off the table. Longer context is still true (Claude defaults 1M tokens, GPT-5.5 defaults 400K, GPT-5.6 defaults 400K standard). And your integration is commodity: ChatGPT Work and Claude Cowork both exist now.
That means the value you bring is now purely about understanding the problem. Do you know what workflows actually need agentic reasoning vs. just retrieval? Do you know which vendor's safety practices matter for regulated industries? Do you know how to architect for model swappability so your client isn't locked in?
Those questions require domain knowledge: about your client's business, not about the model.
The honest counter-take
Claude still wins on some dimensions. Opus 4.8 hasn't been dethroned on SWE-bench Verified (88.6% for Claude, 88.7% for GPT-5.5, 88.8% for GPT-5.6 Sol). That's a point, not a category. On writing quality (instruction-following, long-document analysis, summarization), Claude's outputs are still sharper than GPT's. That's real. Clients doing AI-powered writing (legal docs, content generation, code documentation) notice.
And OpenAI's integration with productivity apps cuts both ways. It's powerful for users who live in Docs and Sheets. It's a liability for enterprises that run custom workflows and don't want their reasoning embedded in Google's infrastructure.
But those are niches now, not value props.
What you should do
Audit your active clients who are using Claude. For each one, ask: What problem are we actually solving? Is it a cost problem, a quality problem, a speed problem, or an integration problem?
- Cost problem: Claude's advantage dies September 1. Migrate those clients to GPT-5.6 Terra, show them the cost parity, and reposition yourself as "I make sure your vendor choice doesn't lock you in."
- Quality problem: Claude still wins on writing. Keep those clients on Claude. Market it as "we're using the model that writes better, not cheaper."
- Speed problem: GPT-5.6 Sol is faster on reasoning. Show benchmarks. Let the client benchmark both. Whoever wins on latency, use that.
- Integration problem: ChatGPT Work solves this for many clients. Some will switch. Others will stay on Claude for the reasons above. Let them choose based on their workflow, not on cost.
The unsellable positioning is "it's cheaper and integrates." That was real for six weeks. It's not real anymore.
Your value is now "I understand your workflow better than your procurement team does." Build that value, or lose the client to ChatGPT Work's simplicity.
Author
Lukas
@lukcombinator