Grok vs GPT midterm study: 1,240 hidden API tokens per call
A Convovo study forced Grok 4.7 and GPT-6.1 to pick midterm candidates. The partisan split grabs headlines; the hidden API tokens matter more to builders.
A research studio led by Cornell professor Tapan Parikh asked two frontier models who to vote for in nearly 500 US midterm races. Per TechCrunch, Grok 4.7 picked the Republican in 92% of answers and GPT-6.1 Sol picked the Democrat in 97%. The politics will get the clicks. The line we care about is quieter: every Grok API call in the Grok vs GPT study carried about 1,240 tokens of input the researchers did not write and could not see.
What actually happened
The study comes from Convovo's AI Ballot 2026 project. It ran through the developer APIs with search on, using a prompt that told each model to name one candidate and not hedge. The consumer apps refused the same question, though Parikh told TechCrunch simple prompts got around that.
Key results, per TechCrunch:
- The two models agreed in less than 10% of races.
- With search off on a 30-race sample, GPT still picked the Democrat in 24 races. Grok declined to pick in 27.
- Grok's answers leaned on different sources than GPT's, which suggests retrieval drove much of Grok's lean.
- Each Grok API request carried roughly 1,240 extra input tokens. Parikh called them potentially "hidden instructions."
OpenAI said the test used a developer API with a prompt that demanded a recommendation, and that ChatGPT "is designed to be objective by default." TechCrunch did not report a response from Grok's maker on the token claim, so we treat it as the researchers' finding, not a confirmed fact.
Why it matters for your business
Your prompt is not the whole prompt. If a provider injects instructions you cannot see, your output depends on text you never reviewed and cannot version. Log input token counts per call and compare them to what you sent. A gap is a question for your vendor.
Search changes the answer. The study suggests the same model behaves differently with retrieval on and off. If your support bot or product assistant browses the web, the sources it pulls shape the reply as much as the model does.
Forced-choice prompts expose defaults. "Pick one, don't hedge" is exactly how many business prompts are written: pick a category, pick a product, approve or reject. Run your own batch of those against two models before a customer sees the result.
Key takeaways
- Convovo found Grok 4.7 picked Republicans in 92% of races and GPT-6.1 Sol picked Democrats in 97%, via developer APIs
- Researchers report about 1,240 unseen input tokens on every Grok API call
- Turning search off changed Grok's behavior far more than GPT's
- Log billed input tokens against what you actually sent
- Test forced-choice prompts on more than one model before production
Putting a model in front of customers? We build AI workflows with logged prompts, token audits and an eval set you own, so a vendor's defaults never become your answer. See how we build or tell us what you're shipping.
Sources: TechCrunch, Convovo AI Ballot 2026.
- #llm-evaluation
- #grok
- #openai
- #api
- #model-bias
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Sweden blocks ChatGPT and Claude for staff; Copilot stays
Sweden's Government Offices blocks ChatGPT and Claude on staff computers from October 26 and keeps Copilot. The issue is the contract, not the model.
Read itJohns Hopkins Bayview AI tool leak: give staff an approved tool
Johns Hopkins Bayview is notifying 3,182 patients after a worker uploaded patient data to a public AI tool. The fix for shadow AI is an approved tool, not a memo.
Read it