Courageous re-ran its personal answer-quality benchmark in June 2026 and positioned its chatbot first, forward of Grok, Google AI Mode, ChatGPT and Perplexity. The corporate revealed the outcomes on June 25 and promoted them once more in the present day by its firm LinkedIn account. The scores exist to promote an API priced at $5 per 1,000 requests.

Brave has up to date the weblog publish saying its revamped Search API with a second spherical of head-to-head efficiency figures, and the revision adjustments the story the corporate can inform about its personal product. Within the first analysis, performed on November 30, 2025, Courageous’s chatbot completed second. Within the June 2026 re-run, it completed first.

The unique publish, titled Brave launches most powerful search API for AI to date, was revealed on February 12, 2026. A word added to the highest of the web page states that the publish was up to date on June 25, 2026 to incorporate scores from the most recent spherical of efficiency comparability exams. Courageous amplified these figures by a publish on its LinkedIn web page, which carries 60,875 followers, drawing 27 reactions, three feedback and 7 reposts on the time of assessment.

The declare being made is slender and particular. In response to Courageous, an open-weights mannequin with higher grounding knowledge can outperform bigger proprietary fashions working on weaker context. That argument is the industrial rationale for the endpoint the corporate is promoting.

What the June analysis reported

Courageous states that the June take a look at repeated the strategy utilized in November. Solutions had been collected from Ask Courageous, Grok, Google AI Mode, ChatGPT and Perplexity throughout the identical set of 1,500 queries randomly sampled from real-world utilization. The weblog publish affords one instance question: “does iPhone accumulate knowledge when off”. Other than Ask Courageous, all solutions had been gathered utilizing BrightData, a scraping service. In response to Courageous, the purpose of utilizing a scraper was to seize the expertise an nameless consumer would truly obtain from every supplier.

Judging was dealt with by massive language fashions moderately than folks. Each Claude Opus 4.5 and Claude Sonnet 4.5 had been utilized in a majority vote setting. Each potential pair of solutions was in contrast, and every pair was evaluated twice, in each orders, to regulate for place bias.

The June outcomes had been introduced on two scales. On a Bradley-Terry Elo score, Ask Courageous scored 1169 and Grok 1157, a niche of 12 factors. The earlier model of Ask Courageous, listed individually as Ask Courageous v1, scored 1098. Google AI Mode sat at 1000, ChatGPT at 880 and Perplexity at 695.

The second column in Courageous’s revealed desk runs 5.0, 4.0, 3.0, 2.0, 1.0 and 0.0 for the six entrants in descending order. The desk header describes that column as an absolute class score on a five-point Likert scale. The bar chart printed straight beneath that desk, on the identical web page, labels the similar values as a Copeland rating, subtitled head-to-head dominance, plotted out of 5 and outlined on the axis as opponents overwhelmed. The weblog’s personal physique textual content describes Copeland as a measure of direct head-to-head comparisons.

Hostinger

The 2 labels can not each be appropriate. Completely spaced integers descending from 5.0 to 0.0 throughout six entrants are the signature of a Copeland rating, not of an averaged high quality score, and the chart resolves the paradox in opposition to the desk header. The excellence adjustments what the underside of the desk means. A Copeland worth of 0.0 for Perplexity doesn’t point out that its solutions had been rated nugatory on a top quality scale. It information that Perplexity misplaced all 5 of its pairwise matchups.

The charts introduce a naming inconsistency of their very own. Each bar charts determine the successful entrant as Ask Courageous v2, whereas the desk above them lists it merely as Ask Courageous alongside a separate Ask Courageous v1. Courageous doesn’t outline anyplace within the publish what separates the 2 variations, how far aside they had been launched, or whether or not v1 corresponds to the configuration that competed in November 2025.

How the outcomes are drawn

The presentation is price studying alongside the numbers. The Elo chart is captioned with a 95% confidence interval, however the bars carry no seen error bars and the publish publishes no interval values, so the acknowledged confidence can’t be checked in opposition to something on the web page. Its horizontal axis begins at 650 moderately than zero, which is a defensible conference for Elo rankings and likewise has the impact of widening the obvious distance between entrants; the 12-point hole separating Ask Courageous from Grok occupies a visibly bigger share of a truncated axis than of a full one.

The November outcomes are charted too, as three side-by-side panels reproducing absolutely the class score, the common win price and the common lose price from the desk. These panels add no figures past what the desk already carries.

How the primary spherical differed

The November 30, 2025 analysis used a real five-point score alongside win and lose charges, and it produced a special rating. Grok led with 4.71, a 59.87% common win price and a ten.05% common lose price. Ask Courageous adopted 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 unfold between first and final place on that scale was 0.70 factors.

Courageous’s personal commentary on the primary spherical conceded the end result, noting that Grok scored finest with Ask Courageous following carefully behind. The corporate attributed its second-place end to the grounding endpoint moderately than to the mannequin, writing that “context high quality issues greater than mannequin functionality.”

Neither spherical discloses the sampling body intimately or states whether or not the June queries had been the similar 1,500 strings utilized in November. The instance question cited adjustments barely between the 2 sections, rendered as “does iphone accumulate knowledge when off” within the first and “does iPhone accumulate knowledge when off” within the second.

Contained in the LLM Context endpoint

The technical substance of the discharge sits in how Courageous assembles the textual content it fingers to a mannequin. Normal net search is constructed round URLs and ranked for human readers. The LLM Context API ranks fragments as a substitute of pages.

In response to Courageous, every request runs a normal net search in opposition to the corporate’s impartial index to determine related pages, then fetches and processes these pages in actual time, changing uncooked HTML into what the corporate calls good chunks. That conversion goes past turning a web page into markdown. Courageous lists query-optimized snippet extraction, structured knowledge extraction protecting JSON+LD schemas, itemprops and tables at row-level granularity, specialised dealing with for code context aimed toward technical questions and coding brokers, extraction of discussion board discussions, and YouTube caption dealing with.

A rating system skilled in-house then selects probably the most related fragments and compiles them right into a response formed by parameters the caller units, together with the whole token depend and the variety of URLs represented.

Hostinger

Latency figures are revealed. Courageous studies lower than 130 milliseconds of overhead on the ninetieth percentile on prime of a traditional search, producing whole latency below 600 milliseconds at p90 for LLM Context calls.

Three controls sit on prime of the endpoint. Goggles, Courageous’s rule-based filtering system, can enhance, downrank or exclude outcomes by area or URL sample and scales to hundreds of guidelines. A token finances parameter, maximum_number_of_tokens, units an higher sure on the estimated measurement of the returned context, with the choice course of prioritizing relevance inside that ceiling. Location headers unlock native outcomes, returning point-of-interest knowledge for companies and map outcomes with geographic context, with the token finances allotted throughout native, map and basic net outcomes.

Courageous states that the endpoint already runs at scale internally, powering over 22 million solutions per day in Courageous Search. The corporate describes that deployment as the most important personal user-facing AI utility on the planet.

The pricing reset

The industrial construction was simplified into 4 public plans: Search, Solutions, Spellcheck and Autocomplete.

Search bundles net outcomes, LLM Context, pictures, information and video below a single price of $5 per 1,000 requests. Solutions, which returns a grounded response alongside the net outcomes that produced it, is priced at $4 per 1,000 net searches plus $5 per million tokens throughout enter and output mixed. Spellcheck and Autocomplete every value $5 per 10,000 requests, with Autocomplete together with entity recognition.

Each plan carries $5 of free credit score that renews month-to-month, which Courageous calculates as $20 throughout all 4. That credit score is conditional. In response to Courageous, builders should attribute the Courageous Search API on their undertaking web site or about web page to assert it.

Courageous positions Solutions and LLM Context at totally different consumers. Solutions fits groups that desire a completed response with minimal setup and studies a 94.1% F1 rating on the SimpleQA benchmark. LLM Context fits groups that need to maintain management of the mannequin layer and provide it with their very own selection of grounding knowledge. The 94.1% determine sits alongside the 94.9% accuracy rating Courageous beforehand cited for its grounding expertise when it launched Ask Courageous, although the 2 are totally different measurements.

Developer distribution and the agent surge

The discharge added two developer-facing items. Courageous Search API Abilities bundle directions, scripts and assets in a standardized, open-sourced format that an AI editor or command-line device can load on demand. Courageous names Cursor, OpenCode and ClaudeCode among the many environments the place these abilities can be utilized. An API Assistant constructed into the Developer Portal solutions questions on endpoints and returns code examples.

The distribution numbers connected to that tooling level to the place demand is coming from. Courageous states that greater than 200,000 builders just lately signed up for its API by the discharge of OpenClaw, the open-source autonomous agent that has unfold quickly since late 2025. A separate Courageous publish dated April 1, 2026 put the OpenClaw-linked determine at near 700,000 customers. Brokers that act on the open net want a search backend, and that requirement has change into a distribution channel for index house owners.

Not each observer is comfy with the underlying class. The Dutch knowledge safety authority warned on February 12, 2026 that open-source brokers of this sort carry materials dangers of information breaches and account takeovers, a warning that applies to the deployment layer moderately than to any single search supplier.

Hostinger

Independence as a gross sales argument

Courageous frames its index as one in all solely three impartial global-scale search indexes within the western world and the one one exterior Massive Tech. That declare carries weight as a result of the alternate options have narrowed. Microsoft shut down its Bing Search APIs on August 11, 2025, pushing builders towards Grounding with Bing Search. Courageous’s personal index has grown from over 20 billion pages in April 2024 to greater than 35 billion pages by the point it launched Ask Courageous, with month-to-month question quantity passing 1.6 billion because the browser reached 101 million month-to-month energetic customers.

The remainder of the pitch is authorized. Courageous argues that scraper-based suppliers violate phrases of service, can not provide true zero knowledge retention, might even see knowledge feeds shut off with out warning, and face latency issues. The weblog publish names Google v. SerpAPI straight as the instance of that publicity, and states that “anybody supplying enterprise options ought to keep away from exposing purchasers to the inherent liabilities of scrapers.”

That case has a documented document. Google sued SerpApi on December 19, 2025 below the Digital Millennium Copyright Act, alleging circumvention of the SearchGuard bot-detection system and in search of $200 to $2,500 per act. SerpApi filed a movement to dismiss on February 20, 2026, arguing amongst different issues that Google lacks standing. Reddit had already sued SerpApi, Oxylabs, AWMProxy and Perplexity on October 22, 2025, and SerpApi moved to kill that amended criticism in March 2026. The litigation is unresolved.

There may be friction between that argument and the benchmark that helps it. Courageous collected each competitor reply in each analysis rounds utilizing BrightData, a scraping service, whereas promoting an API on the premise that scraping carries authorized and operational liabilities its prospects ought to keep away from.

Courageous additionally states that it doesn’t use buyer search queries to coach its personal fashions, affords zero knowledge retention throughout all endpoints, and holds SOC 2 Kind II attestation.

Why this issues for entrepreneurs

The quick relevance just isn’t the benchmark. It’s the value line and the management floor beneath it.

Advertising groups now purchase AI visibility tooling, run generative engine optimisation audits, and more and more construct inner brokers that want dwell net knowledge. Each a type of workloads consumes search API calls. A broadcast price of $5 per 1,000 requests, with token-level finances management on the context payload, offers procurement a set reference level in a market the place Microsoft has retired its public search API and the place scraper-based intermediaries are defending themselves in federal court docket.

The Goggles integration is the much less apparent piece. A group that may enhance or exclude particular domains from the grounding context provided to its personal brokers features management over which sources form inner evaluation. That could be a totally different functionality from optimising to be cited by another person’s assistant, which is the self-discipline that has absorbed most consideration this yr. Semrush analysis protecting 89,000 LinkedIn URLs discovered that quotation behaviour diverges sharply between ChatGPT Search, Google AI Mode and Perplexity, whereas a survey of 45 research discovered that rewriting pages for AI retrieval can minimize a web page’s retrieval price by 16%. Manufacturers are spending in opposition to a floor they don’t management.

Courageous’s framing of the market deserves scrutiny moderately than acceptance. In response to the corporate, “as LLMs change into commoditized, the standard of context they obtain turns into the first differentiator in utility high quality.” Courageous additionally argues that the significance of top quality knowledge ought to shift investor assessments of what drives industrial worth in AI as massive fashions proceed to commoditize. Each statements occur to explain a world wherein Courageous’s asset, an index, appreciates and its opponents’ asset, a mannequin, depreciates.

The impartial verification is lacking. A vendor designed the question set, chosen the entrants, ran the judging pipeline, selected the scoring technique and revealed the outcomes. The judges had been industrial fashions from a 3rd occasion, which provides distance however not neutrality, and the second spherical produced a first-place end the place the primary spherical didn’t. The revealed artefacts don’t absolutely agree with one another both: a desk and the chart beneath it give two totally different names to the identical column, and two totally different names to the identical entrant.

Hostinger

What the tables set up is what Courageous measured. What they don’t set up is how a purchaser’s personal workload would rating.

Timeline

Abstract

Who: Courageous Software program, developer of the Courageous browser and Courageous Search, revealed the outcomes. The benchmark positioned its chatbot Ask Courageous in opposition to Grok, Google AI Mode, ChatGPT and Perplexity. Judging was carried out by Claude Opus 4.5 and Claude Sonnet 4.5.

What: A second spherical of answer-quality comparability scores, added to the weblog publish saying Courageous’s revamped Search API and introduced as each a desk and two bar charts. Ask Courageous positioned first on a Bradley-Terry Elo score at 1169 in opposition to Grok’s 1157, reversing the November 2025 end result. The desk and the chart beneath it label the identical column in another way, as a five-point Likert score and as a Copeland rating. The identical publish particulars the LLM Context endpoint, latency below 600 milliseconds at p90, Goggles and token-budget controls, and pricing at $5 per 1,000 requests on the Search plan.

When: The unique publish was revealed on February 12, 2026 and up to date with the brand new scores on June 25, 2026. The primary analysis ran on November 30, 2025 and the second in June 2026. Courageous promoted the figures on LinkedIn in the present day.

The place: Courageous’s firm weblog below the Courageous Search information part, and the corporate’s LinkedIn web page. The API is bought by the Courageous Search Developer Portal.

Why: Courageous is arguing that grounding knowledge high quality, not mannequin measurement, determines reply high quality, and that an open-weights mannequin provided with higher context can beat frontier programs. The argument helps the endpoint the corporate sells. It arrives as Microsoft has withdrawn its public search API and scraper-based suppliers face DMCA litigation from Google and Reddit, narrowing the variety of impartial net knowledge sources out there to consumers of AI infrastructure.


Source link
Namecheap Best Deals