AI Spar asks your question to multiple AI models simultaneously, finds where they disagree, runs structured debate rounds between them, then produces a balanced synthesis. You get a richer, more trustworthy answer than any single model can give.
Click the ☰ icon (top right on mobile) or look to the right-hand sidebar on desktop to access settings. The lower sections of the panel are collapsed by default — click any section header to expand it.
How many debate rounds to run per disagreement (1–5). More rounds give models more chances to rebut each other but cost more and take longer. 2 rounds is a good default.
When enabled, each model can search the web before answering. Useful for questions about recent events, current prices, or anything that may have changed since a model's training cutoff. Adds a small per-query fee to each model's cost.
Targeted verification is available only for signed-in Grounded debates. It uses the verifier selected in the sidebar to investigate qualifying narrow current factual inputs—such as a fixture, a published probability, a ranking, or odds—when sources are non-inspectable, secondary, conflicting, or stale. It prefers first-party sources and never adjudicates a future prediction. No verifier call is made when no claim qualifies, which is normal.
When it runs, targeted verification adds incremental verifier-model and web-search cost. Its returned material is evidence supplied to the debate, not a final verdict. When a grounded result explicitly links a cited secondary article to a cited primary article, the verifier can make one bounded follow-up to that exact article. It never guesses publisher pages or treats a successful fetch as proof. Debate-visible verification status, outcome, finding, and confidence labels follow the question's language (for example, Chinese questions see Chinese labels). Operator, admin, API, and stored evidence semantics remain canonical. If the core evidence is inspectable but a material qualifier or that primary check remains unresolved, the displayed finding may be mostly supportive with limited confidence; the canonical verdict remains inconclusive.
The AI models that answer your question and debate each other. You can add up to 5 debaters across any combination of providers (Anthropic, OpenAI, Gemini) and models. Click + Add Debater to add one, or the ✕ on a card to remove it.
A separate AI model whose job is to compare debater answers, identify disagreements, and write the final synthesis. It may also check whether a debater has conceded a point (see “What does concession mean?” in the FAQ). The Analyzer is not one of the debaters.
When you are signed in, the model picker refreshes from /api/models/{provider} using the effective inference key, including your Anthropic BYOK key when one is saved. OpenAI choices are gpt-6-luna, gpt-6.1-sol, and gpt-6-astra. Anthropic choices are claude-haiku-5-5, claude-sonnet-5-5, claude-opus-5-5, and claude-fable-5-1; claude-sonnet-5-5 is the default Anthropic model, and a saved claude-sonnet-5 (retired) is replaced with it. Gemini includes gemini-3.5-flash-lite, gemini-3.8-flash, and gemini-4-argon. A saved gemini-3.1-pro-preview (retired) is replaced with gemini-4-argon. gpt-6.1-sol is the default OpenAI model and the default verifier. A saved gpt-6-sol (retired) is replaced with gpt-6.1-sol. Luna and claude-haiku-5-5 (the anonymous Anthropic default) are available to anonymous users. A saved claude-haiku-4-5-20251001 (retired) is replaced with claude-haiku-5-5, for anonymous users too. Astra is hidden until the effective OpenAI key advertises it. Anthropic models unavailable to the effective key are also omitted. Discovery results are filtered to priced, non-retired models.
gpt-6-luna and gpt-6.1-sol for OpenAI, plus the Anthropic choices above. A saved retired or unavailable model is replaced with its successor or the provider default.If you have your own API keys for Anthropic, OpenAI, or Google Gemini, you can store them in AI Spar. When set, your key is used for that provider's models — we do not charge you for those models, and they don't count toward the system fee either. Models from providers where you haven't set a key are charged at actual cost plus the system fee. If you have API keys for all providers in the debate, only the $0.03 base system fee is charged.
To add a key: sign in, open the Settings panel, expand AI Provider API Keys, click Edit next to the provider, paste your key, and click Save. Your keys are encrypted before being stored and never transmitted to the browser.
Click 📎 Attach below the question box to attach images (JPEG, PNG, GIF, WebP) or PDF documents. Up to 3 files, 20 MB each. All debaters and the Analyzer will receive the attachment alongside your question.
Attachments are stored temporarily for 1 hour — enough to complete a debate and any follow-ups before they are deleted.
You can try AI Spar without an account — anonymous users get 2 free questions using the cheapest models from each of the three providers (Anthropic, OpenAI, and Google).
Sign in with Google to unlock:
AI Spar charges after each question completes, not upfront. The charge is:
$0.03 + 25% of non-BYOK token cost. If you bring your own keys for every provider in the debate, the fee is just the $0.03 base.A cost estimate is shown below the question box before you click Ask. The estimate uses typical token averages — actual cost varies with question length and complexity.
Click Add Funds in the sidebar to top up your balance via Stripe. Minimum top-up is $5. Funds do not expire.
Token rates below are per 1 million tokens. Input and output tokens are charged at their respective rates; cached input uses the cache-read rate. Anthropic cache creation uses the cache-write rate shown below. OpenAI writes each prompt to its cache by default and bills the written tokens at the cache-write rate shown below, 1.25× input for the default 30-minute lifetime; a 12-hour cache write bills at 2.5× input (gpt-6-astra $25.00, gpt-6.1-sol $5.00, gpt-6-luna $0.25).
| Model | Provider | Input | Output | Cache read | Cache write |
|---|---|---|---|---|---|
claude-opus-5-5 | Anthropic | $4.00 | $20.00 | $0.20 | $5.00 |
claude-sonnet-5-5 | Anthropic | $2.00 | $10.00 | $0.10 | $2.50 |
claude-haiku-5-5 | Anthropic | $0.10 | $0.50 | $0.01 | $0.125 |
gemini-3.5-flash-lite | $0.30 | $2.50 | $0.03 | — | |
gemini-3.8-flash | $0.75 | $3.75 | $0.075 | — | |
gemini-4-argon | $2.00 | $10.00 | $0.10 | — | |
claude-fable-5-1 | Anthropic | $10.00 | $50.00 | $0.25 | $12.50 |
gpt-6-astra | OpenAI | $10.00 | $50.00 | $1.00 | $12.50 |
gpt-6.1-sol | OpenAI | $2.00 | $10.00 | $0.10 | $2.50 |
gpt-6-luna | OpenAI | $0.10 | $0.50 | $0.01 | $0.125 |
| Model | Provider | Approx. cost per debate* |
|---|---|---|
| gpt-6-luna | OpenAI | $0.04–0.10 |
| gemini-3.5-flash-lite | $0.05–0.25 | |
| claude-haiku-5-5 | Anthropic | $0.04–0.10 |
| gpt-6.1-sol | OpenAI | $0.12–1.00 |
| gemini-3.8-flash | $0.12–0.50 | |
| claude-sonnet-5-5 | Anthropic | $0.20–0.80 |
| gemini-4-argon | $0.70–2.20 | |
| claude-opus-5-5 | Anthropic | $0.80–2.40 |
| claude-fable-5-1 | Anthropic | $2.00–6.00 |
| gpt-6-astra | OpenAI | $0.50–5.00 |
gemini-3.8-flash is promotional at $0.75 input / $3.75 output (cache read $0.075) through December 31, 2026. From January 1, 2027, its rates are $1.50 input / $7.50 output (cache read $0.15); the dated database rows and static estimator schedules change together.
gemini-4-argon is promotional at $2.00 input / $10.00 output (cache read $0.10) through November 30, 2026. From December 1, 2026, its rates are $4.00 input / $20.00 output (cache read $0.20), double the promotional price; its per-debate cost above is at the promotional rate.
Long-context requests bill the whole request at a higher rate. For prompts of more than 272K input tokens: gpt-6.1-sol $4.00 input / $15.00 output (cache read $0.20, cache write $5.00, 12-hour cache write $10.00), gpt-6-sol $4.00 input / $15.00 output (cache read $0.40, cache write $5.00, 12-hour cache write $10.00), gpt-6-luna $0.20 input / $0.75 output (cache read $0.02, cache write $0.25, 12-hour cache write $0.50) and gpt-6-astra $20.00 input / $75.00 output (cache read $2.00, cache write $25.00, 12-hour cache write $50.00). For prompts of more than 200k tokens: gemini-3.1-pro-preview $4.00 input / $18.00 output (cache read $0.40). For prompts of more than 100,000 tokens, counting cached and cache-write tokens: claude-haiku-5-5 $0.50 input / $2.50 output (cache read $0.05, cache write $0.625).
* Per-model estimate for a 2-round debate. Includes all rounds and the system fee. Varies with question length. Grounding adds extra. Legacy models retained in historical conversation records still use their recorded pricing, but are not selectable.
When signed in, every completed debate is automatically saved. Click History in the sidebar to browse past conversations. Click any conversation to reload it and ask follow-up questions.
To bookmark, rename, export, share, or delete a conversation, use the ⋮ menu (or right-click on desktop / long-press on mobile). Use the Bookmarked filter to find saved debates.
Click ⋮ → Share link on any saved conversation. AI Spar generates a permanent public URL (e.g. aispar.app/share/…) and copies it to your clipboard. Anyone with the link can view the full question, synthesis, and debate details — no sign-in required. Sharing is read-only; viewers cannot interact with or continue the conversation.
AI Spar can be used programmatically — from scripts, pipelines, or the command line — using personal access tokens. No browser required.
Sign in, open the Settings panel, scroll to AI Spar API Tokens, enter a label (e.g. my-script), and click Generate. Copy the token immediately — it is only shown once.
aispar CLIThe CLI is a single Python 3 script with no dependencies. Download it once and run it anywhere.
# Download
curl -O https://aispar.app/aispar && chmod +x aispar
# Set your token (add to ~/.bashrc or ~/.zshrc to persist)
export AISPAR_TOKEN=aispar_...All commands output JSON by default (suitable for scripts and pipelines). Add --text for human-readable output.
ask and estimate use the signed-in web defaults: the analyzer is openai:gpt-6-luna, and the missed-point clarification step and multi-point debate are both on. Use --analyzer to pick another analyzer, --no-clarify to skip the clarification step, and --single-point for single-point debate, where each disagreement is debated in its own separate debate instead of all points together in each round (the server still decides how many disagreements are debated).
ask — Run a multi-LLM debate and return the synthesis
# Ask a question — outputs JSON by default
./aispar ask "Is nuclear power safe?"
# Pipe the synthesis into jq
./aispar ask "Is nuclear power safe?" | jq .synthesis
# Human-readable streaming output
./aispar ask "Is nuclear power safe?" --text
# More rounds, custom models
./aispar ask "Is nuclear power safe?" --rounds 3 \
--providers gemini:gemini-3.8-flash anthropic:claude-sonnet-5-5 \
--analyzer openai:gpt-6.1-sol
# Skip clarification and debate each disagreement separately
./aispar ask "Is nuclear power safe?" --no-clarify --single-pointJSON response from ask:
{
"question": "Is nuclear power safe?",
"synthesis": "## Summary\n...",
"charged_usd": 0.382,
"model_cost_usd": 0.32,
"system_fee_usd": 0.062,
"rounds_completed": 2
}balance — Show your current account balance
./aispar balance
./aispar balance --text # Balance: $1.10estimate — Estimate cost before running a debate
./aispar estimate --text
./aispar estimate --rounds 3 \
--providers gemini:gemini-3.8-flash anthropic:claude-sonnet-5-5 \
--analyzer openai:gpt-6-luna \
--texthistory — List your saved conversations
./aispar history --text
./aispar history --limit 5 --textconversation — Show rounds and synthesis for a specific conversation
# Get the ID from `aispar history`
./aispar conversation <id> --texttokens — Manage personal access tokens
Manage tokens from the signed-in web application. Tokens expire automatically; copy the secret when it is created because it is shown only once. Ordinary tokens cannot access administrator endpoints.
Pass your token as a Bearer header. The debate endpoint streams Server-Sent Events — set Accept: text/event-stream.
# Check your balance
curl https://aispar.app/api/user/balance \
-H "Authorization: Bearer $AISPAR_TOKEN"
# Run a debate with the signed-in web defaults (streams SSE events)
curl -N "https://aispar.app/api/debate" \
-H "Authorization: Bearer $AISPAR_TOKEN" \
-H "Accept: text/event-stream" \
-G \
--data-urlencode "question=Is nuclear power safe?" \
--data-urlencode 'providers=[{"type":"gemini","model":"gemini-3.8-flash","name":"gemini gemini-3.8-flash"},{"type":"anthropic","model":"claude-sonnet-5-5","name":"anthropic claude-sonnet-5-5"},{"type":"openai","model":"gpt-6.1-sol","name":"openai gpt-6.1-sol"}]' \
--data-urlencode 'analyzer={"type":"openai","model":"gpt-6-luna","name":"openai gpt-6-luna"}' \
--data-urlencode "clarify_missed_points=true" \
--data-urlencode "multi_point_debate=true" \
--data-urlencode "rounds=2"| Endpoint | Method | Description |
|---|---|---|
/api/debate | GET | Run a debate. Streams SSE. Requires Accept: text/event-stream. |
/api/user/balance | GET | Returns your current balance in USD. |
/api/user/tokens | GET / POST / DELETE | List, create, or revoke access tokens. |
/api/conversations | GET | List your saved conversations. |
/api/conversations/{id} | GET | Fetch a specific conversation with all rounds. |
/api/conversations/{id}/share | POST | Generate (or retrieve) a public share token for a conversation. |
/api/share/{token} | GET | Fetch a shared conversation by token. No authentication required. |
/api/estimate-cost | POST | Estimate cost before running a debate. |
/api/estimate-cost/typical | GET | Typical cost of one debate with the default settings. No authentication required. |
Authorization: Bearer aispar_... or X-Api-Key: aispar_.... Tokens expire automatically and ordinary tokens do not have administrator authority — treat them like passwords and do not share them.
On questions with a well-established consensus answer, models trained on similar data tend to give similar responses. AI Spar will note the consensus and skip the debate phase. This is useful information — it signals the question is not genuinely contested among current AI models.
A concession happens when one model, after reading the other's rebuttal, acknowledges that the other argument is stronger and updates its position. An automated LLM check, which may be made by the Analyzer, can decide whether a model conceded. When the Analyzer's check times out, returns an error or gives a reply that cannot be read, no concession is usually recorded for that response, in single-point and multi-point debates alike. On this service, some of these checks (whether a model conceded a point, which argument persuaded it, and whether it agreed with a point it had initially missed) are made by TypeSafe (typesafe.ai), a separate judgement service, rather than the Analyzer. If TypeSafe returns an error or an answer that cannot be read, the Analyzer usually makes the check instead; for a check of whether a model conceded a point, the fallback described above is used only if the Analyzer also gives no usable answer. If a check of whether a model conceded a point cannot start in time, the fallback described above applies without an Analyzer call. A detected concession can end the debate on a point early. When at least two participants on a point are still responding and all of them have conceded, the point can be recorded as consensus. A point can also end unresolved, for example when fewer than two participants on it are still responding or the rounds run out.
Debaters answer your question and argue positions. The Analyzer never takes a position — its job is to compare answers, extract points of disagreement, and write the final synthesis; it may also judge whether rebuttals constitute concessions. Using a separate model for this role reduces bias and avoids the debaters grading their own work.
Your questions are sent to the AI provider APIs you select (Anthropic, OpenAI, Google). For some automated debate checks, excerpts of the models' answers, which may reflect your question, may also be sent to TypeSafe; see the Privacy Policy. Each provider's own data and privacy policies apply to the content you send them. AI Spar stores your question, the model responses, and the synthesis in its database, associated with your Google account if you're signed in. API keys you store via BYOK are encrypted before being saved.
Before a debate starts, AI Spar checks that your balance can cover it. Normally that means the expected cost fits your balance plus a small overdraft: $1.00 after any top-up, $0.25 before your first top-up, and $0 for suspended accounts. A debate that costs more than expected can take your balance below zero, down to that limit; the next top-up settles it. A negative balance blocks new paid debates until you top up.
If a debate can’t start, the note under the estimate tells you which check it failed. Top up your balance with Add Funds, choose cheaper models, reduce the number of rounds, or add BYOK keys for the expensive providers.
Yes — just type your question in any language. All major models handle Chinese, Japanese, Spanish, French, German, and many others well. The debate and synthesis will typically be in the same language as your question.
If a model fails to respond within the timeout window (120 seconds by default), it is excluded from that round and a note is added to the debate record. If enough models succeed to continue, the debate proceeds. You are not charged for a debate that fails entirely before producing a result.
If you're signed in, completed debates are saved automatically and accessible from the History panel. If you closed the tab while a debate was still running, the debate continues on the server — reload the page and click the same conversation in History to reconnect and see the results when they're ready.
These are the model identifiers used by each provider's API. As a rough guide:
Sign up at each provider's developer platform:
Each provider offers a free tier or trial credits. Once you have a key, paste it into the API Keys section of the AI Spar settings panel.
Have a question, found a bug, or want to share feedback? Send us a message and we'll get back to you.