A model, an assistant, an API, and a gateway are not interchangeable.
Searches for “AI models” often mix underlying neural networks with consumer applications and developer services. That produces weak comparisons: a model benchmark is placed next to an app feature, or an API price is treated as evidence of reasoning quality. Start by naming the layer.
| Layer | What it contains | What to compare |
|---|---|---|
| Foundation or base model | Trained parameters, architecture, tokenizer, and learned capabilities. | Version, provenance, license/access, task evaluations, language coverage, context behavior. |
| Assistant product | User interface, system instructions, tools, memory, uploads, search, policy, and account features. | Actual workflow, privacy, refusal behavior, plan limits, reliability, accessibility. |
| Developer API | Authentication, request schemas, streaming, errors, rate limits, metering, and support. | Supported fields, SDK fit, latency, uptime evidence, cost, observability, data handling. |
| Gateway or model router | A stable product/API layer that can select or wrap configured upstream models. | Routing rules, disclosed identity, consistency, policy boundary, failover, markup, added cost. |
BrokenGPT currently sits primarily in the assistant, API, and gateway layers. It does not claim that the configured Phase 1 foundation model was trained by BrokenGPT.
What the best-known names refer to.
ChatGPT is an assistant product from OpenAI. Claude names Anthropic’s model family and assistant experiences. DeepSeek publishes model and API offerings. Each can change models, plans, features, policies, and limits, so a comparison should always include a date and the exact surface tested.
ChatGPT
Compare the selected OpenAI model and the ChatGPT plan/features separately; the product name alone is not a model version.
Claude
Name the exact Claude model and whether the test used Anthropic’s app, API, or another host.
DeepSeek
Separate the published model revision from DeepSeek’s hosted API and from third-party deployments of released weights.
BrokenGPT
Compare its configured public model, lower-refusal behavior, chat/API path, policy, metering, and disclosure—not an invented base-model claim.
PRIMARY SOURCES
- 01ChatGPT overview
OpenAI — Official product overview.
- 02Claude overview
Anthropic — Official product and model-family overview.
- 03API models and pricing
DeepSeek — Official hosted API documentation; details can change.
Choose with seven questions your workload can answer.
- Task quality: does the exact version solve your real coding, writing, analysis, extraction, or domain task?
- Language quality: does it work in the languages, scripts, dialects, and code-switching your users produce?
- Context behavior: can it retrieve and reason over the input length you need—not merely accept it?
- Tools and modality: do you need search, files, images, audio, code execution, functions, or structured output?
- Behavior and policy: does the product engage with your lawful subject area, and is its boundary compatible with the application?
- Data and control: where are prompts processed, what is retained, can you deploy weights, and which terms govern output?
- Operations and cost: what do input, output, caching, tools, retries, latency, rate limits, and human review cost together?
A public leaderboard can help select candidates, but it cannot replace a workload test. Consumer plans and API products can also behave differently even when they share a model family because tools, system instructions, and serving settings change the result.
BrokenGPT fits a narrower alternative search.
BrokenGPT is relevant when a user wants fewer unnecessary refusals on lawful questions, or when a developer wants a documented chat-completions request shape with bearer keys, JSON/SSE responses, usage-based credits, and separate chat-versus-API accounting. The public model identity, context, serving state, and limitations belong in its model disclosure.
It is not currently positioned as a feature-for-feature ChatGPT or Claude replacement. The public API does not advertise every tool, multimodal, file, agent, or structured-output surface available elsewhere. If those features are mandatory, choose a product that documents them today.
| Priority | Fit today | What to verify |
|---|---|---|
| Lower-refusal lawful chat | Core product direction | Your own prompt set, acceptable-use boundary, factual quality. |
| Server-side text chat API | Documented chat-completions subset | Fields, streaming, errors, limits, billing, SDK version. |
| Transparent usage | Chat and API attribution with one credit balance | Current pricing, metering accuracy, alerts and reporting needs. |
| Multimodal agent platform | Not the advertised public surface | Do not infer files, vision, audio, tools, or agents from text compatibility. |
| Self-hosted foundation-model weights | BrokenGPT is a hosted product layer | Use an upstream open-weight release and its license instead. |
Run a small, dated evaluation before making a big claim.
Build a prompt set from actual user work. Include easy cases, edge cases, long input, multilingual input, prompts likely to be refused, adversarial phrasing, and tasks with independently checkable answers. Run the same set against the exact product plan or API model you are considering.
Record enough context to make the comparison interpretable:
- provider, product surface, model version, date, system instructions, tools, and sampling settings;
- answer correctness, completeness, citations, refusal/deflection, security, and calibrated uncertainty;
- time to first token, completion latency, input/output tokens, errors, retries, and effective cost;
- language/script, reviewer qualifications, scoring rubric, disagreements, and known exclusions.