Researchers at Bitcoin.com News conducted an experiment on October 23, 2026, comparing how three artificial intelligence systems perform when asked to predict bitcoin's closing price on December 31, 2026. The test examined ChatGPT and Grok through voice interfaces, and Claude Opus 5.5 through both text and voice modes, with bitcoin trading near $81,200 at the time of the experiment.
The final price predictions ranged from $78,000 to $85,000 across the four sessions. ChatGPT and Grok voice both settled on $82,000, while Claude's text session predicted $78,000 and its voice session predicted $85,000. However, none of the systems provided transparent calculations demonstrating how their assumptions led to their final figures.
ChatGPT Voice: Fluent but Lacking Demonstrated Analysis
ChatGPT's voice session produced conversational responses that sounded substantive but showed shifting predictions without apparent new research. The forecast moved from $89,500 to $86,500 after additional prompts were introduced, then to $82,000 when longer-term outlook questions were posed. Phrases such as "let me pull those pieces together" and "Checking that now" suggested information retrieval, but the experiment found no evidence such consultation actually occurred. A subsequent text-based ChatGPT conversation offered a more skeptical assessment, admitting it had "treated plausibility as evidence."
Grok Voice: Detailed but Factually Problematic
Grok's spoken responses provided specific figures on bitcoin's historical declines, inflation, institutional activity, and recent developments like the Mt Gox repayment delay. However, the system assigned approximately 90% probability to a Federal Reserve rate hike at an October 27-28 meeting, when market references indicated odds around 6% to 18%. The system also referenced a stale $150,000 Bernstein bitcoin target and incorrectly described Fidelity's projected price range. Grok eventually corrected its rate probabilities when challenged, but correct and incorrect financial claims arrived with similar conversational confidence.
Claude Text Session: Web Search Advantage with Limitations
Claude Opus 5.5's text session performed four web searches before answering and anchored its $78,000 prediction partly to bitcoin's True Market Mean, estimated near $77,600. This provided a more concrete market reference than the ChatGPT voice session offered. However, Claude misstated historical halving-cycle timing, understated bitcoin's peak-to-trough decline, and omitted relevant economic factors from its discussion. The system described using weighted data points and pricing in rate assumptions but later acknowledged it had not computed numerical weights or probabilities. It required three rounds of questioning before adequately separating sourced information from unsupported assumptions.
Claude Voice Model: Most Transparent About Limitations
Claude's voice session provided the experiment's most notable finding. Its initial prediction of $88,000 revised to $85,000, which it maintained through subsequent prompts without unnecessary adjustments. Importantly, Claude volunteered limitations earlier than other sessions, acknowledging it could not pull live onchain metrics and describing its prediction as an estimate rather than a calculated model. When asked to identify unsupported information, it separated retrieved figures, general knowledge, calculations, and opinions on the first request—a disclosure standard that required three rounds of questioning in its own text session.
Claude's voice model also correctly described historical Bitcoin cycle peaks as generally occurring 12 to 18 months after halvings, avoiding timing errors made by both Grok's voice session and Claude's text response. However, the voice session cited outdated oil prices, overstated institutional bitcoin purchases, and omitted a deeper market decline when discussing recent drawdowns. Like the other systems, it never demonstrated a calculation producing its final $85,000 target.
Voice and Text Systems May Differ Significantly
The experiment found that voice and text interfaces for the same AI company may use different underlying systems. OpenAI's ChatGPT voice uses a different model than its text interface, while Claude offers multiple supported versions. These differences complicate straightforward comparisons, as voice experiences may involve different execution settings, latency constraints, and speech-processing components. Grok's voice architecture was not independently verified. The experiment could not establish whether voice AI is inherently less accurate or more prone to fabrication than text-based systems.
Key Takeaway: Appearance of Analysis Differs from Demonstrated Analysis
Across all four sessions, voice assistants delivered financial commentary with professional rhythm and assurance, while text models assembled detailed explanations supported by retrieved sources. Yet all could produce unsupported conclusions. The research concluded that for investors using AI to evaluate financial decisions, rigorous assessment should require identical prompts across sessions, documented retrieval activity, repeated trials, and independently verified figures to meaningfully separate voice interface effects from underlying model reasoning capabilities.
