Armağan Amcalar built a coding school in Berlin before he built a reasoning framework, and the arithmetic behind BRAID carries that engineering instinct. Scope the problem. Measure everything. Ship the thing that survives a spreadsheet. Amcalar is chief technology officer at OpenServ Labs1. Eyup Cinar is a computer engineer at Eskisehir Osmangazi University who also teaches for NVIDIA's Deep Learning Institute. They posted BRAID to arXiv on Dec. 17, 2025. The name stands for Bounded Reasoning for Autonomous Inference and Decisions. Their claim: a flowchart can replace the long monologue a reasoning model writes to itself, and the swap measured up to seventy-four times more work per dollar spent.
The graph replaces the monologue
Reasoning models talk to themselves at length. GPT-5 and its peers write long chains of plain-language thinking before they answer, and every sentence costs tokens, money and time. Amcalar and Cinar state the mechanism directly in the paper. “BRAID introduces a bounded reasoning framework using Mermaid-based instruction graphs that enable models to reason structurally rather than through unbounded natural-language token expansion.”1 The diagram is fixed, machine-readable, and it splits the model that plans from the model that works. The authors call what they replace “reasoning drift”. Loose chains wander, repeat themselves, and sometimes argue against their own premises before landing.
A bounded graph holds its path. Every node carries a set role1. The route from question to answer stays readable enough that OpenServ sells it as “proof of reasoning,” a record that ships with the answer rather than a hidden scratchpad2. Finance, governance and health care top the target list, because those fields already require a record. Tim Hafner, OpenServ's chief executive, puts the payoff in operational terms. “In financial workflows with pricing, allocation, and risk balancing, BRAID maintained consistency where standard models diverged,” he told Benzinga. In a regulated pipeline, consistency beats raw intelligence. The paper's design appears built around that.
What the benchmarks actually show
Three tests anchor the paper1. GSM-Hard, a harder version of the GSM8K math set, ran 100 questions. SCALE MultiChallenge ran 272. AdvancedIF ran 100 more, built around hard instructions. Each test pairs two roles BRAID assigns apart: a generator that drafts the Mermaid graph, and a solver that runs it. The authors tried pairings across five GPT tiers, from gpt-4o through the gpt-5 medium, mini and nano variants, several at “minimal” effort. A costly model can plan while a cheap one works, or the roles can flip.
Results followed. A gpt-5-nano-minimal solver lifted GSM-Hard accuracy from 94.0% to 98.0%1. On SCALE MultiChallenge, gpt-4o climbed from 19.9% to 53.7%, nearly triple, while gpt-5-nano-minimal rose from 23.9% to 45.2%. AdvancedIF accuracy for gpt-5-nano-minimal more than doubled, from 18.0% to 40.0%.
BRAID nearly tripled gpt-4o's SCALE MultiChallenge accuracy, from 19.9% to 53.7%
Accuracy on GSM-Hard, SCALE MultiChallenge and AdvancedIF, percent of questions answered correctly, at baseline and with BRAID
Each row is one model on one test: the blue dot is its baseline accuracy, the magenta dot is its accuracy with BRAID, and the rod is the gain.
Source: arXiv [1]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Baseline | With BRAID | Note |
|---|---|---|---|
| MultiChallenge, gpt-4o | 19.9% | 53.7% | Nearly triple its baseline |
| AdvancedIF, gpt-5-nano-minimal | 18.0% | 40.0% | More than double its baseline |
| MultiChallenge, gpt-5-nano-minimal | 23.9% | 45.2% | |
| GSM-Hard, gpt-5-nano-minimal | 94.0% | 98.0% |
Accuracy alone undersells the case, so the authors built a second metric1. Performance-per-dollar sets a gpt-5-medium baseline at 1.0. A gpt-4.1 generator paired with a gpt-5-nano-minimal solver on GSM-Hard scored 74.06, at 96% accuracy against the baseline's 95%. That is where the headline “74x” comes from. AdvancedIF's best pairing scored 61.69, and SCALE MultiChallenge's best scored 30.31. Benzinga reported a fourth number from the plain GSM8K set2. GPT-5 scored 64.34 with BRAID, against 54.41 under the classic setup, with a claimed 25% to 40% cut in cost per correct answer in financial work.
BRAID's best GSM-Hard pairing scored 74.06 on performance per dollar, against a gpt-5-medium baseline of 1.0
Performance-per-dollar score of the best BRAID pairing on each benchmark, indexed to a gpt-5-medium baseline of 1.0
Each bar is one score on the paper's performance-per-dollar index: the magenta bar is the 74x headline pairing, gray bars are the best pairings on two other tests, and the blue bar is the baseline.
Source: arXiv [1]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value | Note |
|---|---|---|
| GSM-Hard best pairing | 74.06 | A gpt-4.1 generator with a gpt-5-nano-minimal solver, at 96% accuracy against the baseline's 95% |
| AdvancedIF best pairing | 61.69 | |
| MultiChallenge best pairing | 30.31 | |
| gpt-5-medium baseline | 1.00 | The paper's reference point, set at 1.0 |
Caveats ride along with all of it. The authors say savings assume Mermaid graphs get cached and reused, and a fresh graph per query costs much more1. GSM-Hard needed a “Numerical Masking Protocol” to stop solvers from reading answers already sitting in the diagram nodes. That hints at how easily this kind of test leaks its own answers. Grading used GPT-5.2 as an automated judge, looser than exact matching. Every model compared came from OpenAI, a limit the paper sets on purpose, leaving Anthropic, Google and open-weight models for later.
Pricing reasoning by the dollar it saves
Chain-of-thought pricing rewards length by design. Vendors meter tokens, and a model that thinks at length makes more of them. DeepSeek's reasoning API shows the range, $0.14 to $0.87 per million tokens as of September 20268. That spread is wide enough that one design choice inside a single vendor's lineup can rival the gap between vendors. BRAID's metric turns the spread into a dial a team can turn.
Reuse sits at the center of the math. The caveat, that savings assume a graph gets cached and used across many similar queries, doubles as a business model. Value goes to whoever runs the cache, rather than to whoever calls the API once. That explains why OpenServ sells BRAID as infrastructure, a layer for repeat work like pricing, allocation and risk, where one graph fires thousands of times a day.
Skeptics will note the incentive. SERV, the company's token, trades on exactly this story. The economics still hold past any one vendor. Buyers weighing agent infrastructure now ask what a correct answer costs, alongside what a benchmark says, and dollars per correct output speaks to that.
The proof threshold
Scrutiny arrived within months. CryptoSlate's Liam “Akiba” Wright, under the headline “the real test starts now,” argued the benchmark claims were doing the work while real proof stayed out of reach3. “If the result depends on bounded tasks with narrow decision trees, that can still be useful in enterprise settings where reliability and cost control carry more weight than a broad conversational range,” Wright wrote. A buyer running BRAID against its own workload meets that exact fork. Wright's standard rests on three questions. Which models were compared, under which conditions? Did tasks come from public benchmarks or internal composites? How much of any cost advantage traces to model choice rather than orchestration? OpenServ has marketed its lightweight SERV Nano model as able to “match or beat OpenAI” at “20x lower cost and 3x the speed,” language Wright flagged as promotional and still undocumented by outsiders.
Markets moved before critics finished reading. SERV rallied roughly 70% on the news, per Bitget and BeInCrypto, briefly pushing toward a $39 million valuation before easing back45. CryptoSlate put the market value in the mid-teens of millions by the time its piece ran3. A token that moves 70% on a research paper is its own data point. Very little now separates a benchmark from a trade.
Who is behind the graph
Amcalar's path runs through teaching before research. He founded Coyotiv, a School of Software Engineering in Berlin, and built a public reputation under the GitHub handle dashersw6. Entrepreneur's UK edition covered the Coyotiv and OpenServ work on cutting reasoning costs, reading his move into research as an extension of the teaching. Cinar supplies the academic side, an engineering post at Eskisehir Osmangazi University plus instructor credentials at NVIDIA's Deep Learning Institute12.
Outside interest followed fast. Bahadır Akdemir published a Medium explainer on how structured reasoning makes a model sharper and cheaper, which shows the argument traveled past OpenServ's own channels7. Coyotiv keeps working as Amcalar's base6. The school sells itself as a next-generation engineering ecosystem, and its graduates feed the talent OpenServ can draw on as the team grows past two authors.
Hafner handles enterprise sales, and Amcalar handles the research2. Both face the harder job now. Turning a strong paper and a token chart into deployments that regulators and finance chiefs approve takes more than a benchmark. Cinar's dual role, half academic and half NVIDIA instructor, puts BRAID in the university review pipeline and on the practitioner circuit at once. That matters while the paper's boldest claim, seventy-four times the work per dollar, waits on replication outside the lab that made it.
By the numbers
- A score of 74.06 on performance-per-dollar came from a gpt-4.1 generator paired with a gpt-5-nano-minimal solver, against a baseline of 1.01.
- SCALE MultiChallenge accuracy for gpt-4o under BRAID reached 53.7%, nearly triple its 19.9% baseline1.
- AdvancedIF accuracy for gpt-5-nano-minimal reached 40%, more than double an 18.0% baseline1.
- GPT-5 scored 64.34 on GSM8K with BRAID applied, against 54.41 under the classic setup, per Benzinga2.
- Cost per correct answer fell 25% to 40% across financial workflow testing, by OpenServ's account2.
- SERV rallied roughly 70% after the benchmark claim spread, then cooled into a smaller-cap range45.
- Dec. 17, 2025: the day Amcalar and Cinar submitted BRAID to arXiv, months before the market noticed1.
- Between $0.14 and $0.87 per million tokens: DeepSeek's reasoning API pricing as of September 20268.
What to watch
Replication decides whether the gap survives engineers outside OpenServ's payroll, and open Mermaid implementations already circulating give researchers a way to test it1. Watch whether OpenAI, Anthropic or DeepSeek publish bounded-reasoning work of their own, because a real cost edge rarely stays private once buyers ask vendors to match it. Pilots in finance and health care, the two fields OpenServ names, should give the clearest signal2. Auditability earns its premium the moment a regulator asks to see the record.
Sources
- Armağan Amcalar and Eyup Cinar, "BRAID: Bounded Reasoning for Autonomous Inference and Decisions," arXiv, Dec. 17, 2025, https://arxiv.org/abs/2512.15959.
- Murtuza J Merchant, "OpenServ's BRAID Framework Surpasses GPT Models, Targets Enterprise Use With Auditable AI Reasoning," Benzinga, Aug. 28, 2025, https://www.benzinga.com/crypto/cryptocurrency/25/08/47387496/openservs-braid-framework-surpasses-gpt-models-targets-enterprise-use-with-auditable-ai-reasoning.
- Liam "Akiba" Wright, "Crypto AI project OpenServ says it can beat OpenAI, but the real test starts now," CryptoSlate, Apr. 6, 2026, https://cryptoslate.com/openserv-openai-benchmark-claims-proof-threshold/.
- "SERV Surges 70% After AI Benchmark Claim: Can a $39M Token Really Beat GPT-5.4?" Bitget News, Apr. 2026, https://www.bitget.com/news/detail/12560605415736.
- "This Altcoin Soars 70% on AI Agent Hype: Why The Rally Could Cool Fast," BeInCrypto, Apr. 2026, https://beincrypto.com/openserv-serv-falling-wedge-breakout-ai-agents/.
- "Coyotiv and OpenServ Are Working to Cut AI Reasoning Costs," Entrepreneur (UK edition), 2026, https://uk.entrepreneur.com/technology/coyotiv-and-openserv-are-working-to-cut-ai-reasoning-costs/503898.
- Bahadır Akdemir, "BRAID: How Structured Reasoning Can Make Your LLM Smarter and Cheaper," Medium, 2026, https://medium.com/@akdemir_bahadir/braid-how-structured-reasoning-can-make-your-llm-smarter-and-cheaper-c90045537b91.
- "DeepSeek API Pricing (September 2026): $0.14–$0.87 per 1M Tokens," BenchLM.ai, September 2026, https://benchlm.ai/deepseek/api-pricing.