Brave re-ran its own answer-quality benchmark in June 2026 and placed its chatbot first, ahead of Grok, Google AI Mode, ChatGPT and Perplexity. The company published the results on June 25 and promoted them again today through its company LinkedIn account. The scores exist to sell an API priced at $5 per 1,000 requests.

Brave has updated the blog post announcing its revamped Search API with a second round of head-to-head performance figures, and the revision changes the story the company can tell about its own product. In the first evaluation, conducted on November 30, 2025, Brave's chatbot finished second. In the June 2026 re-run, it finished first.

The original post, titled Brave launches most powerful search API for AI to date, was published on February 12, 2026. A note added to the top of the page states that the post was updated on June 25, 2026 to include scores from the latest round of performance comparison tests. Brave amplified those figures through a post on its LinkedIn page, which carries 60,875 followers, drawing 27 reactions, three comments and seven reposts at the time of review.

The claim being made is narrow and specific. According to Brave, an open-weights model with better grounding data can outperform larger proprietary models running on weaker context. That argument is the commercial rationale for the endpoint the company is selling.

What the June evaluation reported

Brave states that the June test repeated the method used in November. Answers were collected from Ask Brave, Grok, Google AI Mode, ChatGPT and Perplexity across the same set of 1,500 queries randomly sampled from real-world usage. The blog post offers one example query: "does iPhone collect data when off". Apart from Ask Brave, all answers were gathered using BrightData, a scraping service. According to Brave, the point of using a scraper was to capture the experience an anonymous user would actually receive from each provider.

Judging was handled by large language models rather than people. Both Claude Opus 4.5 and Claude Sonnet 4.5 were used in a majority vote setting. Every possible pair of answers was compared, and each pair was evaluated twice, in both orders, to control for position bias.

The June results were presented on two scales. On a Bradley-Terry Elo rating, Ask Brave scored 1169 and Grok 1157, a gap of 12 points. The previous version of Ask Brave, listed separately as Ask Brave v1, scored 1098. Google AI Mode sat at 1000, ChatGPT at 880 and Perplexity at 695.

The second column in Brave's published table runs 5.0, 4.0, 3.0, 2.0, 1.0 and 0.0 for the six entrants in descending order. The table header describes that column as an absolute category rating on a five-point Likert scale. The bar chart printed directly beneath that table, on the same page, labels the identical values as a Copeland score, subtitled head-to-head dominance, plotted out of five and defined on the axis as opponents beaten. The blog's own body text describes Copeland as a measure of direct head-to-head comparisons.

The two labels cannot both be correct. Perfectly spaced integers descending from 5.0 to 0.0 across six entrants are the signature of a Copeland ranking, not of an averaged quality rating, and the chart resolves the ambiguity against the table header. The distinction changes what the bottom of the table means. A Copeland value of 0.0 for Perplexity does not indicate that its answers were rated worthless on a quality scale. It records that Perplexity lost all five of its pairwise matchups.

The charts introduce a naming inconsistency of their own. Both bar charts identify the winning entrant as Ask Brave v2, while the table above them lists it simply as Ask Brave alongside a separate Ask Brave v1. Brave does not define anywhere in the post what separates the two versions, how far apart they were released, or whether v1 corresponds to the configuration that competed in November 2025.

How the results are drawn

The presentation is worth reading alongside the numbers. The Elo chart is captioned with a 95% confidence interval, but the bars carry no visible error bars and the post publishes no interval values, so the stated confidence cannot be checked against anything on the page. Its horizontal axis begins at 650 rather than zero, which is a defensible convention for Elo ratings and also has the effect of widening the apparent distance between entrants; the 12-point gap separating Ask Brave from Grok occupies a visibly larger share of a truncated axis than of a full one.

The November results are charted too, as three side-by-side panels reproducing the absolute category rating, the average win rate and the average lose rate from the table. Those panels add no figures beyond what the table already carries.

How the first round differed

The November 30, 2025 evaluation used a genuine five-point rating alongside win and lose rates, and it produced a different ranking. Grok led with 4.71, a 59.87% average win rate and a 10.05% average lose rate. Ask Brave followed at 4.66, with 49.21% wins and 15.82% losses. Google AI Mode recorded 4.39, ChatGPT 4.32 and Perplexity 4.01. The spread between first and last place on that scale was 0.70 points.

Brave's own commentary on the first round conceded the outcome, noting that Grok scored best with Ask Brave following closely behind. The company attributed its second-place finish to the grounding endpoint rather than to the model, writing that "context quality matters more than model capability."

Neither round discloses the sampling frame in detail or states whether the June queries were the identical 1,500 strings used in November. The example query cited changes slightly between the two sections, rendered as "does iphone collect data when off" in the first and "does iPhone collect data when off" in the second.

Inside the LLM Context endpoint

The technical substance of the release sits in how Brave assembles the text it hands to a model. Standard web search is built around URLs and ranked for human readers. The LLM Context API ranks fragments instead of pages.

According to Brave, each request runs a standard web search against the company's independent index to identify relevant pages, then fetches and processes those pages in real time, converting raw HTML into what the company calls smart chunks. That conversion goes beyond turning a page into markdown. Brave lists query-optimized snippet extraction, structured data extraction covering JSON+LD schemas, itemprops and tables at row-level granularity, specialized handling for code context aimed at technical questions and coding agents, extraction of forum discussions, and YouTube caption handling.

A ranking system trained in-house then selects the most relevant fragments and compiles them into a response shaped by parameters the caller sets, including the total token count and the number of URLs represented.

Latency figures are published. Brave reports less than 130 milliseconds of overhead at the 90th percentile on top of a normal search, producing total latency under 600 milliseconds at p90 for LLM Context calls.

Three controls sit on top of the endpoint. Goggles, Brave's rule-based filtering system, can boost, downrank or exclude results by domain or URL pattern and scales to thousands of rules. A token budget parameter, maximum_number_of_tokens, sets an upper bound on the estimated size of the returned context, with the selection process prioritizing relevance within that ceiling. Location headers unlock local results, returning point-of-interest data for businesses and map results with geographic context, with the token budget allocated across local, map and general web results.

Brave states that the endpoint already runs at scale internally, powering over 22 million answers per day in Brave Search. The company describes that deployment as the largest private user-facing AI application in the world.

The pricing reset

The commercial structure was simplified into four public plans: Search, Answers, Spellcheck and Autocomplete.

Search bundles web results, LLM Context, images, news and video under a single rate of $5 per 1,000 requests. Answers, which returns a grounded response alongside the web results that produced it, is priced at $4 per 1,000 web searches plus $5 per million tokens across input and output combined. Spellcheck and Autocomplete each cost $5 per 10,000 requests, with Autocomplete including entity recognition.

Every plan carries $5 of free credit that renews monthly, which Brave calculates as $20 across all four. That credit is conditional. According to Brave, developers must attribute the Brave Search API on their project website or about page to claim it.

Brave positions Answers and LLM Context at different buyers. Answers suits teams that want a finished response with minimal setup and reports a 94.1% F1 score on the SimpleQA benchmark. LLM Context suits teams that want to keep control of the model layer and supply it with their own choice of grounding data. The 94.1% figure sits alongside the 94.9% accuracy score Brave previously cited for its grounding technology when it introduced Ask Brave, though the two are different measurements.

Developer distribution and the agent surge

The release added two developer-facing pieces. Brave Search API Skills package instructions, scripts and resources in a standardized, open-sourced format that an AI editor or command-line tool can load on demand. Brave names Cursor, OpenCode and ClaudeCode among the environments where those skills can be used. An API Assistant built into the Developer Portal answers questions about endpoints and returns code examples.

The distribution numbers attached to that tooling point to where demand is coming from. Brave states that more than 200,000 developers recently signed up for its API through the release of OpenClaw, the open-source autonomous agent that has spread rapidly since late 2025. A separate Brave post dated April 1, 2026 put the OpenClaw-linked figure at close to 700,000 users. Agents that act on the open web need a search backend, and that requirement has become a distribution channel for index owners.

Not every observer is comfortable with the underlying category. The Dutch data protection authority warned on February 12, 2026 that open-source agents of this type carry material risks of data breaches and account takeovers, a caution that applies to the deployment layer rather than to any single search provider.

Independence as a sales argument

Brave frames its index as one of only three independent global-scale search indexes in the western world and the only one outside Big Tech. That claim carries weight because the alternatives have narrowed. Microsoft shut down its Bing Search APIs on August 11, 2025, pushing developers toward Grounding with Bing Search. Brave's own index has grown from over 20 billion pages in April 2024 to more than 35 billion pages by the time it launched Ask Brave, with monthly query volume passing 1.6 billion as the browser reached 101 million monthly active users.

The rest of the pitch is legal. Brave argues that scraper-based providers violate terms of service, cannot offer true zero data retention, may see data feeds shut off without warning, and face latency problems. The blog post names Google v. SerpAPI directly as the example of that exposure, and states that "anyone supplying enterprise solutions should avoid exposing clients to the inherent liabilities of scrapers."

That case has a documented record. Google sued SerpApi on December 19, 2025 under the Digital Millennium Copyright Act, alleging circumvention of the SearchGuard bot-detection system and seeking $200 to $2,500 per act. SerpApi filed a motion to dismiss on February 20, 2026, arguing among other things that Google lacks standing. Reddit had already sued SerpApi, Oxylabs, AWMProxy and Perplexity on October 22, 2025, and SerpApi moved to kill that amended complaint in March 2026. The litigation is unresolved.

There is friction between that argument and the benchmark that supports it. Brave collected every competitor answer in both evaluation rounds using BrightData, a scraping service, while selling an API on the premise that scraping carries legal and operational liabilities its customers should avoid.

Brave also states that it does not use customer search queries to train its own models, offers zero data retention across all endpoints, and holds SOC 2 Type II attestation.

Why this matters for marketers

The immediate relevance is not the benchmark. It is the cost line and the control surface underneath it.

Marketing teams now buy AI visibility tooling, run generative engine optimisation audits, and increasingly build internal agents that need live web data. Every one of those workloads consumes search API calls. A published rate of $5 per 1,000 requests, with token-level budget control on the context payload, gives procurement a fixed reference point in a market where Microsoft has retired its public search API and where scraper-based intermediaries are defending themselves in federal court.

The Goggles integration is the less obvious piece. A team that can boost or exclude specific domains from the grounding context supplied to its own agents gains control over which sources shape internal analysis. That is a different capability from optimising to be cited by someone else's assistant, which is the discipline that has absorbed most attention this year. Semrush research covering 89,000 LinkedIn URLs found that citation behaviour diverges sharply between ChatGPT Search, Google AI Mode and Perplexity, while a survey of 45 studies found that rewriting pages for AI retrieval can cut a page's retrieval rate by 16%. Brands are spending against a surface they do not control.

Brave's framing of the market deserves scrutiny rather than acceptance. According to the company, "as LLMs become commoditized, the quality of context they receive becomes the primary differentiator in application quality." Brave also argues that the importance of high quality data should shift investor assessments of what drives commercial value in AI as large models continue to commoditize. Both statements happen to describe a world in which Brave's asset, an index, appreciates and its competitors' asset, a model, depreciates.

The independent verification is missing. A vendor designed the query set, selected the entrants, ran the judging pipeline, chose the scoring method and published the results. The judges were commercial models from a third party, which adds distance but not neutrality, and the second round produced a first-place finish where the first round did not. The published artefacts do not fully agree with each other either: a table and the chart beneath it give two different names to the same column, and two different names to the same entrant.

What the tables establish is what Brave measured. What they do not establish is how a buyer's own workload would score.

Timeline

Summary

Who: Brave Software, developer of the Brave browser and Brave Search, published the results. The benchmark placed its chatbot Ask Brave against Grok, Google AI Mode, ChatGPT and Perplexity. Judging was performed by Claude Opus 4.5 and Claude Sonnet 4.5.

What: A second round of answer-quality comparison scores, added to the blog post announcing Brave's revamped Search API and presented as both a table and two bar charts. Ask Brave placed first on a Bradley-Terry Elo rating at 1169 against Grok's 1157, reversing the November 2025 result. The table and the chart beneath it label the same column differently, as a five-point Likert rating and as a Copeland score. The same post details the LLM Context endpoint, latency under 600 milliseconds at p90, Goggles and token-budget controls, and pricing at $5 per 1,000 requests on the Search plan.

When: The original post was published on February 12, 2026 and updated with the new scores on June 25, 2026. The first evaluation ran on November 30, 2025 and the second in June 2026. Brave promoted the figures on LinkedIn today.

Where: Brave's company blog under the Brave Search news section, and the company's LinkedIn page. The API is sold through the Brave Search Developer Portal.

Why: Brave is arguing that grounding data quality, not model size, determines answer quality, and that an open-weights model supplied with better context can beat frontier systems. The argument supports the endpoint the company sells. It arrives as Microsoft has withdrawn its public search API and scraper-based providers face DMCA litigation from Google and Reddit, narrowing the number of independent web data sources available to buyers of AI infrastructure.