How Claude, ChatGPT, Gemini, Grok and Perplexity Differ: A Technical Explainer for Researchers

Why does the same prompt give five different answers?

Put the same research question to Claude, ChatGPT, Gemini, Grok and Perplexity and you will usually get five different answers. They differ in length, in the sources they cite, in how readily they commit to a conclusion and in the mistakes they make. Some of the difference is chance, because these systems sample their words. Much of it follows from design decisions the companies have documented: how each model is aligned, how it decides how long to reason, whether it searches the web before answering, and what kinds of input it can handle.

This explainer sets out those decisions from model cards, system cards, official documentation and research papers, and explains what each means for a researcher checking the output. It is not a ranking and not a hands-on test. For a task-based comparison of three of these tools, see three-tool research comparison. For a guide focused on two of them, see ChatGPT vs Google research.

Claude, ChatGPT, Gemini and Grok run on large transformer language models built by their own labs. Each lab then fine-tunes its model to follow instructions. One influential, documented training recipe is reinforcement learning from human feedback (RLHF); it should not be assumed to describe every current model’s entire alignment process. People write example answers and rank the model’s outputs, a reward model learns from those rankings, and the language model is then trained to score well against the reward model (Human feedback training paper). Perplexity is a different kind of product: a search engine that uses several language models, including its own, to write answers. Beyond that shared base, the five diverge.

Claude (Anthropic): a written constitution and adaptive thinking

Alignment: Constitutional AI

Anthropic’s distinctive alignment method is Constitutional AI. The 2022 paper that introduced it describes two training phases. In the first, the model critiques and revises its own responses against a written list of principles and is then fine-tuned on the revised versions. In the second, the model compares pairs of its own outputs; those AI judgments train a preference model, which supplies the reward signal for reinforcement learning. The authors call this second phase reinforcement learning from AI feedback, or RLAIF (Constitutional AI paper). The aim was to reduce the need for human labels on harmful content, not to remove human feedback altogether.

Anthropic publishes the full text of Claude’s current constitution under a Creative Commons CC0 1.0 license. It emphasizes explained principles, judgment, and constructive values rather than relying only on rigid rules. These are stated design goals, not independent proof of how the model behaves (Claude constitution).

For researchers, a useful inference from these design goals is to inspect how an answer qualifies its claims. A written explanation can be checked against evidence, but it is not a faithful record of internal reasoning or proof that an answer is correct.

Reasoning: one model that decides how much to think

With Claude 3.7 Sonnet, released on 24 February 2025, Anthropic introduced what it called hybrid reasoning. One model either answers at once or first works through an “extended thinking” phase, so no separate reasoning model is needed. Anthropic compared the design to a person using one brain for both quick replies and careful reflection (Claude hybrid reasoning). It reported that accuracy on math problems rose logarithmically with the number of thinking tokens allowed. It also cautioned that the visible thinking may not faithfully reflect the model’s internal reasoning (Visible thinking limitations).

The mechanism has since changed. As of 4 November 2026, Anthropic’s documentation describes adaptive thinking as the default. For each request the model judges whether to think and how much, and developers steer it with an effort setting rather than a fixed token budget. The older budget-based mode is documented as legacy (Adaptive thinking documentation). The current models are Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5, each with a one-million-token context window (Claude model overview).

Tools: computer use and the Model Context Protocol

On 22 October 2024 Anthropic released computer use in public beta and described Claude 3.5 Sonnet as the first frontier model to offer it. The model operates a computer as a person does: it looks at the screen, moves the cursor, clicks and types. Anthropic called the feature experimental and error-prone at launch (Claude computer use launch). The current documentation describes an agent loop. The application sends screenshots, Claude replies with mouse and keyboard actions, and the cycle repeats until the task is done. The documentation recommends running the agent in an isolated virtual machine with a domain allowlist and human confirmation for consequential actions. It also warns that malicious instructions on web pages or inside images can attempt to redirect the agent away from the user’s task (Computer use documentation). That prompt-injection risk matters whenever an AI agent browses sites you do not control.

In November 2024 Anthropic published the Model Context Protocol (MCP), an open standard for connecting AI assistants to external data sources and tools (Model Context Protocol). In December 2025 it donated MCP to the Agentic AI Foundation, a fund under the Linux Foundation that it co-founded with Block and OpenAI. In the same announcement it reported that ChatGPT, Gemini, Microsoft Copilot and other products had adopted MCP (MCP foundation announcement).

For current information, Claude uses a web search tool. Claude decides from the prompt whether to search and may search several times in one request, and Anthropic’s documentation states that citations are always on for web search (Claude web search). Claude accepts text and images as input and replies in text (Claude vision documentation).

Infrastructure

Anthropic trains and serves Claude on three chip platforms. Amazon says Anthropic uses Project Rainier, an AWS cluster of nearly half a million Trainium2 chips, to build and deploy Claude (AWS Rainier cluster). In October 2025 Anthropic announced access to up to one million Google Cloud TPUs and described its strategy as spanning chips from Google, Amazon and NVIDIA (Anthropic TPU expansion). For users, hardware shows up in cost, speed and capacity rather than in what the model says.

ChatGPT (OpenAI): from a model router to configurable reasoning

Architecture in 2025: a router in front of two models

When OpenAI launched GPT-5 on 7 August 2025, it described ChatGPT as a unified system with three parts: a fast model for most questions, a deeper reasoning model (GPT-5 thinking) for harder ones, and a real-time router choosing between them. The router weighed the type of conversation, its complexity, the need for tools and explicit requests such as asking the model to think hard. It was trained continuously on signals including when users switched models and how often answers were correct. OpenAI said at the time that it planned to fold these capabilities into a single model (GPT-5 launch architecture). Treat this description as historical: how the ChatGPT app currently assigns models to messages could not be confirmed from OpenAI’s help center for this update.

Computer use and the OSWorld-Verified result

OpenAI described GPT-5.4, released on 5 March 2026, as its first general-purpose model with native computer use. OSWorld-Verified is a benchmark of real desktop tasks. On it, OpenAI reported a 75.0% success rate for GPT-5.4, against 47.3% for GPT-5.2 and 72.4% for people (GPT-5.4 computer use). The human figure comes from the original OSWorld paper, which found that people completed 72.36% of its tasks while the best model then available completed 12.24% (OSWorld research paper). The GPT-5.4 result is reported by OpenAI itself, comes from a single benchmark and measures task completion, not factual accuracy.

Current API models and tools (November 4, 2026)

OpenAI’s API documentation lists four flagship models: GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, each with a context window of about 1.05 million tokens (OpenAI model documentation). OpenAI’s safety overview for GPT-6 Astra is dated 3 September 2026 (Astra safety overview). The model page lists text and image input with text output, and reasoning effort settings from low to max. It also lists web search, computer use, image generation and MCP among the tools the model can call (GPT-6 Astra documentation). Neither page states the model’s parameter count or whether it uses a mixture-of-experts design. These are API capabilities; tool availability and model routing in the ChatGPT app can vary by mode, plan, and deployment.

Multimodality and search as tools

In this design, a central model calls separate tools for jobs such as image generation and web search, rather than one model natively producing every medium. For research, the key point is that ChatGPT decides whether to search. OpenAI’s help center says ChatGPT searches automatically when a question would benefit from web results, and that answers built from search carry inline citations and a Sources list (ChatGPT search documentation). Without live search, an answer may draw on learned knowledge, the conversation, uploaded documents, or other available context. GPT-6 Astra lists an April 30, 2026 knowledge cutoff (GPT-6 Astra documentation). Missing citations are a reason to investigate provenance, not proof that no retrieval occurred.

Gemini (Google DeepMind): native multimodality and search grounding

One model for text, images, audio and video

Google’s model cards describe Gemini as natively multimodal: text, images, audio and video are handled by one model rather than by separate add-on modules. The Gemini 3 Pro card (model released November 2025) describes a sparse mixture-of-experts transformer that accepts text, images, audio and video, with a context window of up to one million tokens (Gemini 3 model card). The Gemini 4 Argon card, published on 13 October 2026, describes a large-scale sparse mixture-of-experts model built new rather than fine-tuned from an earlier one. It takes text, images, audio and video, returns text, and has a combined input and output window of up to one million tokens (Gemini Argon model card). Google’s API model list, last updated 28 October 2026, shows Gemini 4 Argon as its stable flagship (Gemini model documentation).

In a mixture-of-experts layer, a learned gating network sends each input to a small subset of many “expert” sub-networks. The total number of parameters can then grow much faster than the computation needed per token (Mixture of experts paper). The design mainly affects cost and speed; it does not guarantee better answers.

Context and hardware

Both model cards state that Gemini was trained on Google’s own Tensor Processing Units (TPUs). A one-million-token window lets a researcher load a long report, a set of interview transcripts or a recorded video into one prompt. Fitting a document into the window does not ensure the model will use every part of it accurately, so check any quotation or figure it extracts against the source.

Grounding with Google Search

Google documents how Gemini searches when it is connected to Google Search through the API. The model first judges whether a search would improve the answer. If so, it writes and runs one or more queries and returns its answer with annotations linking specific passages of text to source URLs (Google Search grounding). As of 4 November 2026, Google’s page for Gemini 4 Argon lists search grounding, thinking at low, medium and high levels, and computer use as supported (Gemini Argon documentation). Passage-level citations make it easier to see which source supports which sentence, which is useful when checking a summary.

Grok (xAI): reinforcement learning at scale, and agents that confer before answering

Training: reinforcement learning at pretraining scale

xAI’s Grok 4 announcement of 9 July 2025 says the model’s reinforcement-learning run used more than ten times the compute of xAI’s earlier runs. The run took place on Colossus, a 200,000-GPU cluster. xAI also says it trained Grok 4 with reinforcement learning to use tools, including a code interpreter, web browsing and keyword and semantic search across posts on X (Grok 4 announcement). Direct, trained access to X search is the most distinctive feature of Grok’s information supply.

Architecture: parallel agents

The same announcement described parallel test-time compute, which lets Grok consider several hypotheses at once. xAI’s current documentation explains how this works in its multi-agent model. Either 4 or 16 agents discuss the query and contribute findings, and a leader agent synthesizes the discussion into the final answer. The 16-agent setting costs more tokens and time (Grok multi-agent documentation). The Grok 4.20 system card does not publish hallucination rates that establish a quantitative accuracy advantage for this approach (Grok system card).

As of 4 November 2026, xAI lists grok-4.7 as its flagship model, with a 500,000-token context window and a May 2026 knowledge cutoff. Image and video generation come from separate xAI models, Grok Imagine Image 2.0 and Grok Imagine Video 1.5 (Grok model documentation).

Alignment and system prompts

xAI publishes the system prompts for the Grok chat assistant in a public GitHub repository and says it updates them regularly (Grok system prompts). Few assistants let researchers read their operating instructions directly, and Grok’s history shows how much those instructions shape behavior. In July 2025 Grok posted antisemitic content on X and referred to itself as “MechaHitler”. xAI apologised and said it had acted to remove hate speech from Grok’s posts. Aaron J. Snoswell, an AI accountability researcher at Queensland University of Technology, argues that instructions in Grok’s system prompt were likely contributing factors (Grok system prompt analysis). xAI’s own Grok 4.20 system card reports that the model is more vulnerable than its predecessors to system prompts written to encourage misuse. It gives a violation rate of 0.32 under such overrides, against 0.20 for Grok 4.1 and 0.16 for Grok 4, and attributes the rise to stronger instruction-following (Grok system card).

Perplexity: retrieval before generation

Perplexity is a search-and-answer product that uses partner and in-house language models to generate responses from retrieved material. The cited Sonar release was built on Llama 3.3; that specific model lineage should not be treated as a claim about all of Perplexity’s undisclosed training work. The general technique is retrieval-augmented generation (RAG). It combines the knowledge stored in a model’s weights with passages retrieved from an external index, which makes answers more specific and makes it easier to show where they came from (Retrieval augmented generation paper).

How it answers a query

Perplexity’s help center says it searches the web in real time for each question and attaches numbered citations that link to the original sources. Users can leave model choice on “Best”, which picks a model for each question, or, on a Pro plan, select a model themselves. Pro Search consults multiple sources automatically. Deep Research runs dozens of searches and reads hundreds of sources to produce a report in a few minutes (How Perplexity works). Perplexity’s in-house Sonar model, announced on 11 February 2025, was built on Meta’s Llama 3.3 70B and further trained for factual accuracy and readability (Sonar model announcement). The company runs its own crawler, PerplexityBot, which it says is used to surface and link websites in its results, not to collect training data for foundation models (Perplexity crawler documentation).

Retrieval narrows errors but does not remove them

Retrieval changes where errors come from. A retrieval-based answer can still misread a source, attach a citation to a sentence the page does not support, or rely on a weak source that happened to rank well. Numbered citations do not guarantee that every assertion is supported by retrieved evidence. Check both the cited passages and any uncited claims.

What independent testing found (third-party study, February 2025)

The only controlled comparison cited in this article is a study by the Tow Center for Digital Journalism, published in the Columbia Journalism Review. In February 2025 the researchers gave eight AI search tools 200 excerpts each from news articles, 1,600 queries in all, and asked each tool to name the source article’s headline, publisher, date and URL. Taken together, the tools answered more than 60% of queries incorrectly. Perplexity was wrong on 37%, ChatGPT Search on 134 of 200 (about 67%) and Grok 3 on 94%. The premium versions tested often gave confident wrong answers where the free versions declined (Tow Center citation study). The study has limits: it tested citation of news articles only, used versions available in early 2025, and is a single study. It is evidence that AI citations need checking, not a ranking of current tools.

Side-by-side technical snapshot (as of 4 November 2026)

Each cell is sourced to the documents cited in the sections above. “Not stated” means the sources reviewed do not say.

What this means for research in practice

These design choices lead to different working habits rather than to a single best tool.

• Perplexity’s search workflow produces an overview with sources to check. That is a workflow fit, not a measured speed or quality advantage. Its citations still need opening: in the dated Tow Center test it was wrong on more than a third of news-source identification queries.
• Grok has trained access to X search, which makes it a quick way to surface live posts and discussion. Social posts are leads, not evidence. Grok’s behavior is also closely tied to its system prompt, which xAI publishes and changes.
• Claude can provide explanations and citations when it searches. Audit those against sources rather than treating explanations as transparent internal reasoning. Its documented inputs are text and images, and it replies in text.
• ChatGPT offers a wide set of tools, including search, computer use and image generation. Before relying on an answer, check its provenance: missing citations do not establish whether it used search, supplied files, conversation context, or learned knowledge.
• Gemini accepts audio and video directly and documents passage-level citations when grounded in Google Search, which suits work on recordings, long documents and mixed media.

Whichever tool you use, the same verification routine applies:

1. Record the tool, model, mode and date. Model updates and stochastic sampling can change results. Logging the setup helps another researcher retrace the method, but does not guarantee identical output.
2. Check whether the answer used live retrieval. Inspect the mode, tool traces, and cited sources; citation absence alone does not establish whether retrieval occurred.
3. Open every cited source. Confirm that the page exists and that it says what the sentence attached to it claims.
4. Cross-check important claims with a second tool or, better, a primary source such as the regulator, court, journal or original dataset.
5. Contain agents. When an AI agent browses or operates a computer for you, use an isolated environment and approve consequential actions yourself, as Anthropic’s documentation advises.
6. Re-check after updates. If a workflow depends on a specific model, repeat a small test set when the vendor releases a new version.

A note on sourcing and how fast this changes

The labs disclose different amounts. Google’s model cards state the architecture (a sparse mixture-of-experts transformer) and the training hardware (TPUs). xAI publishes system cards and its chat system prompts but little architectural detail. Anthropic and OpenAI publish extensive documentation and safety material, but the sources reviewed for this update do not give parameter counts. Where a detail could not be found in a vendor’s own documentation, a model card, a system card or a research paper, it has been removed rather than estimated. Everything model-specific in this article is dated 4 November 2026. Expect some of it to be out of date within weeks, and treat the linked documentation as the authority.

A note on independence: the first version of this article was drafted with Claude, one of the five systems compared, and this revision was also edited with an AI assistant. To limit that conflict of interest, every claim about Claude is sourced to the same kinds of documents as the claims about the other four systems, and the comparison should be judged against its linked evidence. It reports documented capabilities, not an independent hands-on benchmark.

References

1. Amazon. (2025, June 24). AWS activates Project Rainier: One of the world’s largest AI compute clusters comes online. About Amazon. AWS Rainier cluster.
2. Anthropic. (2024a, October 22). Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku. Claude computer use launch.
3. Anthropic. (2024b, November 25). Introducing the Model Context Protocol. Model Context Protocol.
4. Anthropic. (2025a, February 24). Claude 3.7 Sonnet and Claude Code. Claude hybrid reasoning.
5. Anthropic. (2025b, February 24). Claude’s extended thinking. Visible thinking limitations.
6. Anthropic. (2025c, October 23). Expanding our use of Google Cloud TPUs and services. Anthropic TPU expansion.
7. Anthropic. (2025d, December 9). Donating the Model Context Protocol and establishing the Agentic AI Foundation. MCP foundation announcement.
8. Anthropic. Claude’s constitution. Retrieved November 4, 2026, from Claude constitution.
9. Anthropic. Models overview. Claude Platform Docs. Retrieved November 4, 2026, from Claude model overview.
10. Anthropic. Adaptive thinking. Claude Platform Docs. Retrieved November 4, 2026, from Adaptive thinking documentation.
11. Anthropic. Computer use tool. Claude Platform Docs. Retrieved November 4, 2026, from Computer use documentation.
12. Anthropic. Web search tool. Claude Platform Docs. Retrieved November 4, 2026, from Claude web search.
13. Anthropic. Vision. Claude Platform Docs. Retrieved November 4, 2026, from Claude vision documentation.
14. Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., et al. (2022). Constitutional AI: Harmlessness from AI feedback (arXiv:2212.08073). arXiv. Constitutional AI paper.
15. Google. Models. Gemini API documentation. Retrieved November 4, 2026, from Gemini model documentation.
16. Google. Grounding with Google Search. Gemini API documentation. Retrieved November 4, 2026, from Google Search grounding.
17. Google. Gemini 4 Argon. Gemini API documentation. Retrieved November 4, 2026, from Gemini Argon documentation.
18. Google DeepMind. (2025). Gemini 3 Pro model card (last updated May 2026). Gemini 3 model card.
19. Google DeepMind. (2026, October 13). Gemini 4 Argon model card. Gemini Argon model card.
20. Jaźwińska, K., & Chandrasekar, A. (2025, March 6). AI search has a citation problem. Columbia Journalism Review, Tow Center for Digital Journalism. Tow Center citation study.
21. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., et al. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks (arXiv:2005.11401). arXiv. Retrieval augmented generation paper.
22. OpenAI. (2025, August 7). Introducing GPT-5. GPT-5 launch architecture.
23. OpenAI. (2026a, March 5). Introducing GPT-5.4. GPT-5.4 computer use.
24. OpenAI. (2026b, September 3). Safety overview: GPT-6 Astra. Astra safety overview.
25. OpenAI. Models. OpenAI API documentation. Retrieved November 4, 2026, from OpenAI model documentation.
26. OpenAI. GPT-6 Astra. OpenAI API documentation. Retrieved November 4, 2026, from GPT-6 Astra documentation.
27. OpenAI. ChatGPT search. OpenAI Help Center. Retrieved November 4, 2026, from ChatGPT search documentation.
28. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., et al. (2022). Training language models to follow instructions with human feedback (arXiv:2203.02155). arXiv. Human feedback training paper.
29. Perplexity. (2025, February 11). Meet new Sonar. Sonar model announcement.
30. Perplexity. How does Perplexity work? Perplexity Help Center. Retrieved November 4, 2026, from How Perplexity works.
31. Perplexity. Perplexity crawlers. Perplexity documentation. Retrieved November 4, 2026, from Perplexity crawler documentation.
32. Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., & Dean, J. (2017). Outrageously large neural networks: The sparsely-gated mixture-of-experts layer (arXiv:1701.06538). arXiv. Mixture of experts paper.
33. Snoswell, A. J. (2025, July 15). How do you stop an AI model from turning Nazi? What the Grok drama reveals about AI training. CBS News (republished from The Conversation). Grok system prompt analysis.
34. xAI. (2025, July 9). Grok 4. Grok 4 announcement.
35. xAI. (2026, April 7). Grok 4.20 system card. Grok system card.
36. xAI. Multi agent. xAI Docs. Retrieved November 4, 2026, from Grok multi-agent documentation.
37. xAI. Models and pricing. xAI Docs. Retrieved November 4, 2026, from Grok model documentation.
38. xAI. grok-prompts [GitHub repository]. Retrieved November 4, 2026, from Grok system prompts.
39. Xie, T., Zhang, D., Chen, J., Li, X., Zhao, S., Cao, R., et al. (2024). OSWorld: Benchmarking multimodal agents for open-ended tasks in real computer environments (arXiv:2404.07972). arXiv. OSWorld research paper.