OpenTranscription/ Blog
Subscribe
◆ 2026-07-03 · ANALYSIS

Deepgram Enhanced behind Future AGI's Agent Command Center: what the public record actually shows

A close read of Deepgram's Enhanced STT tier and Future AGI's Agent Command Center gateway, including the documentation gap between the two.

OpenTranscription 2026-07-03
See the model profile →
Abstract illustration of an audio waveform passing through a series of geometric gateway layers before reaching a speech model lattice

Deepgram's Enhanced is one of the stranger objects in the current speech-to-text market: a model tier launched in May 2022, still listed in Deepgram's docs and rate-limit tables, still routable through third-party gateways, and yet almost invisible in Deepgram's own 2025 and 2026 product story. At launch, Deepgram called it a more accurate ASR tier built on its "next generation End-to-End Deep Learning speech model architecture," claiming 19% higher relative accuracy than the previous model and materially better handling of long-tail vocabulary. Today it survives in variants such as general, meeting, phonecall, and finance while the company's public narrative centers on Nova-3, Flux, and newer speech products. Deepgram does not publicly disclose Enhanced's parameter count, architecture diagram, training corpus size, or model-specific WER benchmarks in any of the sources reviewed here.

On the other side of this pairing sits Future AGI's Agent Command Center, a gateway and control-plane layer between an application and its model providers. The public docs describe a fixed plugin pipeline covering authentication, caching, budgets, guardrails, rate limiting, the provider call itself, cost tracking, and logging. It exposes OpenAI-compatible APIs, per-request metadata headers, Prometheus and OpenTelemetry exports, exact and semantic caching, shadow experiments, and self-hosted deployment. The public code and documentation surface names 15 routing and reliability strategies, though only a subset are fully explained in the formal docs; the rest are named in the open-source gateway README and related pages, with some details left unspecified.

And here is the finding that makes this pairing worth writing about at all: Future AGI's public documentation has a gap around Deepgram specifically. Its provider docs list many cloud and self-hosted LLM providers and expose /v1/audio/transcriptions for Whisper-style STT, yet its public model catalog simultaneously advertises routable model IDs such as deepgram/enhanced, deepgram/enhanced-general, deepgram/enhanced-phonecall, deepgram/enhanced-meeting, and deepgram/enhanced-finance, each described as callable through Agent Command Center with unified observability, caching, fallback, and "15 routing strategies including cost-optimized fallback." The integration is publicly marketed and partially documented, but not yet reconciled with the main provider and API docs.

The defensible conclusion for anyone evaluating this stack: Enhanced is real and routable, but public technical transparency around it is thin. Agent Command Center is a credible gateway layer for routing, observability, and fallback, but the exact Deepgram adapter semantics for Enhanced are not fully specified in the formal docs currently available. If you are an engineering buyer or architect, plan on verification work around provider registration, endpoint translation, media handling, and fallback semantics for audio transcription requests before you bet production traffic on it.

The companies and the people behind them

Deepgram was founded in 2015. Its official company story says the business emerged from machine-learning work on waveform analysis in a dark-matter detector in China. Scott Stephenson, identified by Deepgram as CEO and co-founder, later explored deep learning for audio analysis at the University of Michigan before building Deepgram around end-to-end deep learning. Deepgram's author bio for Stephenson states he earned a PhD in particle physics from the University of Michigan and left a postdoctoral research role to found the company.

Future AGI presents itself as an open-source, end-to-end AI agent engineering platform spanning tracing, evaluation, optimization, protection, and gateway functions. Its founder pages are less formalized than Deepgram's, but public profiles and company pages identify Nikhil Pareek as Founder and CEO, and multiple public profiles and startup directories identify Charu Gupta as co-founder. The same public sources indicate Future AGI was founded in 2024, though the official homepage does not spell out that founding year.

The personnel picture is asymmetrical. Deepgram has a clearer official company-history narrative but relatively little founder detail beyond Stephenson on its official pages. Future AGI has a richer scattered public footprint across blog posts, webinar pages, GitHub, LinkedIn snippets, and interviews, but less of it is consolidated into a single formal leadership page.

Company Person Publicly confirmed role Bio details visible in reviewed sources Public profiles Source
Deepgram Scott Stephenson CEO, co-founder Deepgram describes him as a dark-matter physicist turned deep-learning entrepreneur; PhD in particle physics, University of Michigan; left postdoc work to found Deepgram. Deepgram author page; LinkedIn public profile snippet
Deepgram Other co-founders Not fully specified in official pages reviewed Deepgram's official story references "Scott Stephenson and his teammate" but does not name the teammate on the pages reviewed. Unspecified in official material reviewed
Deepgram Andrew Seagraves VP of Research Official Deepgram byline identifies him as VP of Research; additional biography details were not published on the page reviewed. Deepgram article byline
Future AGI Nikhil Pareek Founder & CEO Public profiles describe him as Founder/CEO; Forbes profile says he has about a decade in AI, a prior exit, and this is his second funded startup; a founder interview says his first job involved autonomous agents for drones. LinkedIn; Forbes Council profile; Cerebral Valley interview
Future AGI Charu Gupta Co-founder Public profiles and a founder story describe her as co-founder with 15+ years of business scaling experience. LinkedIn public profile snippet; founder story
Future AGI N.V.J.K. Kartik Founding Engineer Public LinkedIn snippet identifies him as Founding Engineer at Future AGI; education listed as IIIT Dharwad. Official webinar page says he "ships the routing and caching surface" of Agent Command Center. LinkedIn; Future AGI webinar page
Future AGI Rishav Hada Senior Applied Scientist Official Future AGI webinar page says he leads the guardrails segment; multiple Future AGI research/blog pages identify him as Senior Applied Scientist. A personal site describes him as a Mila graduate researcher focused on reliable ML systems. Future AGI blog/webinar pages; personal site
Future AGI Nikita Sklyarov Technical Lead Public LinkedIn snippet describes him as Technical Lead building agentic AI systems; more detailed background was not specified in the reviewed official sources. LinkedIn public snippet

Why Enhanced exists and what Deepgram has actually disclosed

Deepgram's explanation for building newer speech tiers is consistent across its 2022 materials. End-to-end deep learning made STT more flexible than traditional pipelines because Deepgram did not have to separately reconnect and re-optimize acoustic, pronunciation, and language models after each change. That made it easier to retrain and improve models without starting over. Deepgram positioned Enhanced explicitly as the middle tier between Base and custom-trained models: better out-of-the-box quality than Base, without the cost and effort of full custom model training.

The market need shows up clearly in Deepgram's keywording guidance. Some teams cannot justify custom training but still need stronger recognition of uncommon or long-tail terms. For those teams, Enhanced was the recommended step up because it had an increased effective vocabulary and handled infrequent words significantly better than lower tiers. That tradeoff matters for how you read the product. Enhanced was never pitched as universal frontier ASR. It was a practical tier for organizations that need better domain-term recall without committing to a custom-training program.

On the technical side, the strongest official statement remains the launch announcement: Enhanced is "based on our next generation End-to-End Deep Learning speech model architecture," with 19% higher relative accuracy than the previous model, better word recognition, and stronger long-tail vocabulary handling. Deepgram also warned customers that model upgrades can change transcript outputs, and recommended side-by-side testing with features such as keywords before production rollout.

What is not publicly specified is just as important. In the official sources reviewed, Deepgram does not disclose Enhanced's parameter count, its training data size, its architecture family beyond "next generation end-to-end deep learning," its serving topology, latency-to-first-token, official WER on named datasets, or any competitor tables specific to Enhanced. By 2025 and 2026, Deepgram's public benchmark language had moved on to Nova-2, Nova-3, and Flux.

Abstract waveform with sparse tall amber spikes representing rare long-tail vocabulary terms being caught by a fine geometric mesh

Where Enhanced sits in Deepgram's model lineup

Model/tier Official positioning Public architecture/training disclosure Public accuracy disclosure Public options noted Service limits visible in docs Source
Base Cost-effective tier on Deepgram's end-to-end STT architecture Built on Deepgram's "signature end-to-end deep learning" architecture; no detailed model card on reviewed page No headline WER on reviewed page general, meeting, phonecall, voicemail, finance, conversationalai, video Enhanced/Base/Nova families share concurrency tables in current limits docs
Enhanced Higher-accuracy tier above Base "Next generation End-to-End Deep Learning speech model architecture"; no parameter count or dataset card publicly disclosed 19% higher relative accuracy vs prior model; improved long-tail vocabulary general, meeting beta, phonecall, finance beta Starter/Growth/Enterprise concurrency tables list Enhanced for both pre-recorded and streaming
Nova Predecessor to Nova-2 Docs say training spanned 100+ domains and 47B tokens No single WER figure on the model page excerpt reviewed general, phonecall Same concurrency table family
Nova-3 Current flagship general ASR Deepgram's current benchmark and product pages emphasize it; formal architecture details still limited in docs reviewed 6.84% median WER streaming, 5.26% batch; Deepgram claims 54.3% and 47.4% relative reductions vs competitors General-purpose current flagship Current rate limits list Nova-3 prominently
Flux Conversational streaming ASR with turn detection Positioned as latest-generation streaming/conversational model Independent Coval validation cited by Deepgram for latency/turn-taking; not directly an Enhanced successor Conversational/turn-based experiences Current limits list Flux streaming separately

Inputs, outputs, and throughput for Enhanced

Enhanced is exposed through Deepgram's standard STT interfaces rather than a bespoke API. Pre-recorded STT uses the /v1/listen family and accepts audio and video inputs, with JSON responses for transcription results or a request ID for callback and asynchronous paths. Streaming uses WebSockets and supports standard streaming controls plus transcript options such as sample rate, smart formatting, diarization, and interim versus final results, along with the supported audio encodings. Deepgram's docs state support for 100+ audio formats and encodings overall.

Current Deepgram rate-limit docs still list Enhanced for concurrency. On starter-like limits in the North America and Europe tables, Enhanced is shown at up to 50 concurrent pre-recorded requests and up to 150 concurrent streaming requests, rising on Growth and Enterprise plans. Those numbers are service limits, not intrinsic model throughput measurements, but they are the most concrete concurrency figures publicly visible for Enhanced.

Deepgram also makes broader performance claims that give some context, though not specifically for Enhanced. Its 2022 materials described the use-case models as able to transcribe one hour of pre-recorded audio in about 30 seconds, and a 2022 G2-oriented post described Deepgram real-time streaming as having less than 300 ms of lag. Treat both as family-level or platform-level marketing claims rather than Enhanced-only measurements.

Inside Agent Command Center

Future AGI describes Agent Command Center as a single OpenAI-compatible gateway between the client application and upstream providers. The architecture docs say each request passes through a fixed-order plugin chain: IP ACL, authentication, RBAC, cache lookup, budget checks, guardrails, tool policy, validation, and rate limiting before the provider call. The response then continues through cost tracking and logging. Cache hits can short-circuit the provider call entirely.

The integration value of routing Deepgram Enhanced through this gateway is not "making Enhanced exist." It is wrapping Enhanced in a standard control plane: virtual keys instead of raw provider keys, per-call metadata, cache controls, cost headers, routing and fallback behaviors, OpenTelemetry and Prometheus observability, and shadow experiments. Future AGI's public Deepgram model pages say that calling models such as deepgram/enhanced-general through Agent Command Center returns metadata on x-agentcc-* headers including provider, cost, latency, cache hit, and request ID.

There is a caveat worth sitting with. Agent Command Center's provider docs enumerate OpenAI, Anthropic, Gemini, Bedrock, Azure, Cohere, Groq, Mistral, Together, Fireworks, DeepInfra, Perplexity, Cerebras, xAI, OpenRouter, Hugging Face, Anyscale, Replicate, and several self-hosted backends. They do not mention Deepgram. At the same time, Future AGI's Deepgram model catalog advertises 36 Deepgram models, including the Enhanced variants, as routable through Agent Command Center. That likely means one of three things: the public provider docs lag the catalog, the catalog uses an internal or generic adapter path that is not fully documented, or the calculator and catalog surface is ahead of the formal endpoint docs. The public sources reviewed do not resolve which.

The request pipeline described in the architecture docs, combined with the publicly advertised Deepgram catalog integration, gives a coherent picture: plugin order and telemetry and caching semantics come from official docs, while the Deepgram target path comes from Future AGI's Deepgram model pages.

Stylized signal-flow diagram of a request passing through a sequence of abstract geometric gate shapes, with one path short-circuiting back early to represent a cache hit

The 15 routing strategies, and how well each is documented

Future AGI's public README names 15 routing and reliability strategies exactly as follows: roundrobin, latency, costopt, adaptive, complexity, conditional, providerlock, accessgroups, race/hedged, mirror/shadow, modelfallback, failover, circuitbreaker, retry, healthmonitor. The formal routing docs fully explain only a subset. Where the public algorithm is not described, the table below marks it accordingly.

Strategy What the public docs support Detail level in public docs Source
Round robin Evenly rotates traffic across providers; default example strategy Fully specified
Weighted Splits traffic by assigned weights Fully specified
Least latency Routes to fastest provider using recent response times Fully specified
Cost optimized Chooses cheapest provider supporting the requested model Fully specified
Adaptive Dynamically adjusts weights using real-time performance Fully specified at high level
Complexity-based Scores requests on 8 signals and maps them to a model tier Fully specified at high level
Conditional Docs/config mention "conditional routing rules"; exact rule language is not fully documented on the pages reviewed Partially specified
Provider lock README names it; likely forces/pins a request to a chosen provider for compliance or policy reasons, but exact semantics were not publicly documented in reviewed pages Named only
Access groups Public docs explain access groups as logical sets of models/aliases for policy management Partially specified
Race / hedged README names hedged/race requests; implies parallel or near-parallel upstream calls to reduce tail latency, but exact policy knobs were not documented in reviewed pages Named only
Mirror / shadow Shadow experiments mirror a sampled portion of production traffic to a target model/provider without affecting users Well specified
Model fallback Per-model ordered fallback chains; e.g., if one model fails, try alternatives in sequence Fully specified
Failover Triggers on 429, 5xx, timeouts, and connection errors; routes to backup provider list Fully specified
Circuit breaker Opens on repeated failures, then half-opens for recovery probes Fully specified
Retry Exponential backoff with configurable retry counts and backoff windows Fully specified
Health monitor Public provider-health docs describe continuous health tracking plus cooldown/probe-based reentry Partially specified

The routing and fallback flow that emerges from these docs, taking in the plugin order, cache behavior, failover, retry, circuit-breaker logic, and shadow experiments, is a synthesis of public documentation rather than a copied vendor diagram.

Observability, caching, APIs, and deployment

The observability surface is the strongest part of the documentation. Future AGI says every request is logged with request ID, trace ID, session ID, model requested versus model actually used, provider, token counts, cost, latency, cache status, guardrail results, and fallback or error events. The same docs state that the gateway exports metrics to Prometheus and traces to OpenTelemetry, and that users can propagate their own trace IDs via request headers.

Caching is server-side and supports exact and semantic strategies. Public docs describe configurable TTL, namespaces, LRU eviction, force-refresh, and Cache-Control: no-store. The open-source README states support for 6 exact-cache backends (mem, redis, disk, s3, gcs, azblob) and 4 semantic-cache backends (mem, pinecone, qdrant, weaviate).

On APIs and SDKs, the formal docs present Agent Command Center as OpenAI-compatible, with endpoints including /v1/chat/completions, /v1/embeddings, /v1/audio/transcriptions, /v1/audio/speech, /v1/images/generations, and /v1/responses. Future AGI publishes a dedicated client SDK repository for Python and TypeScript, while the control-plane and gateway service itself lives in the broader platform monorepo.

Deployment options are a hosted gateway, Docker, and Go-binary self-hosting. The self-hosted docs say Agent Command Center ships as a Go binary and Docker image, can run from a simple config.yaml, and is intended to keep requests fully inside the customer's infrastructure for data residency and network-topology control. The README adds that clustering and HA use Raft-based clustering, and that official Kubernetes manifests and Helm charts were "coming soon" at the time of the README snapshot reviewed.

For security, the gateway relies on virtual API keys, encrypted storage of provider credentials, RBAC, rate limits, budget controls, IP allowlists, and append-only audit logging per the open-source README. The docs also emphasize that prompts and completions are not stored by default; caching is opt-in and configurable per organization.

Benchmarks, code, and patents

The numbers that exist, and what they actually measure

Scope Figure or claim What it means Caveat Source
Deepgram Enhanced launch 19% higher relative accuracy vs previous model Official launch claim for Enhanced tier No public WER table or named benchmark dataset in reviewed source
Deepgram long-tail vocabulary "Increased effective vocabulary," better handling of uncommon words Product rationale for Enhanced Qualitative claim, not benchmarked in reviewed docs
Deepgram family speed One hour of audio in ~30 seconds; <300 ms real-time lag Historical Deepgram platform-level speed claims Not Enhanced-specific and partly older marketing
Deepgram Nova-3 6.84% median WER streaming; 5.26% batch Current official flagship benchmark Applies to Nova-3, not Enhanced
Future AGI gateway throughput ~28,889 req/s on 4 vCPU / 16 GB t3.xlarge profile Gateway throughput under mock-upstream benchmark Measures gateway overhead, not upstream model latency
Future AGI gateway latency P95 2.8 ms at ~1k RPS; full chat proxy ~66 µs internal wall time Fast gateway/control-plane overhead Self-published benchmark against mock upstream
Independent STT critique 44% average transcription error on spoken U.S. street names across evaluated top-provider models Real-world high-stakes error failure mode Not Enhanced-specific; cross-provider study
Independent accent critique Deepgram and peers vary materially on non-native accented English Highlights robustness gaps beyond headline WER Not Enhanced-specific
Independent audit critique Standard ASR audits can understate weakness for aphasia speakers Quality varies by speech type and evaluation method Not Enhanced-specific

Open-source code and reproducibility

Deepgram publishes official developer SDKs and playground tooling for its STT, TTS, and voice APIs. The docs point to the API Playground, JavaScript SDK, and getting-started guides for pre-recorded and live transcription. That is useful for reproducing functional API behavior, but not for reproducing Enhanced's model training or benchmark internals.

Future AGI's public code surface is more explicit on reproducibility. The future-agi/future-agi monorepo includes the agentcc-gateway README and benchmark harness, while future-agi/agent-command-center-sdk publishes client SDKs. future-agi/traceAI is an Apache 2.0 tracing layer built on OpenTelemetry. The gateway README says benchmark configs, commands, and mock-upstream harnesses are committed under bench/, with go test -bench and bench/run.sh reproduction instructions.

Patents and IP

Deepgram's January 2026 Series C press release says the company expanded its patent portfolio and specifically cites US 12,499,875 for "Deep Learning Internal State Index-Based Search and Classification," framing it as protection for techniques that leverage internal neural representations for large-scale audio search and classification. USPTO patentee indexes also show Deepgram assignee entries for "Hardware efficient automatic speech recognition" (granted June 2025) and "End-to-end automatic speech recognition with transformer" (granted August 2025). These patents matter because they point to Deepgram's twin emphases on inference efficiency and end-to-end ASR modeling, even though they do not amount to a public model card for Enhanced.

Primary sources worth reading directly

Type Item Why it matters Source
Official launch note Deepgram Enhanced launch changelog, May 26 2022 Primary source for Enhanced's 19% relative-accuracy claim and next-gen E2EDL positioning
Official model docs Deepgram model options page Current documentation that Enhanced is still listed, with available options
Official product rationale Deepgram article on model selection and keyword boosting Explains why Enhanced sits between base and custom models, especially for long-tail vocabulary
Official architecture docs Future AGI Command Center Architecture Primary source for request pipeline, plugin order, caching short-circuit, and multi-tenant virtual keys
Official routing docs Future AGI Routing, Failover, and Load Balancing Primary source for round-robin, weighted, least-latency, cost-optimized, adaptive, failover, retries, circuit breaker, and complexity routing
Official observability docs Future AGI Observability, Cost Tracking, Caching, Shadow Experiments Primary sources for unified observability, headers, shadow traffic, exact/semantic cache, and budget/cost telemetry
Open-source gateway README future-agi/future-agi agentcc-gateway/README.md Primary source for the 15-strategy list, benchmark numbers, cache backends, and deployment posture
Official Deepgram history/about Deepgram About + Scott Stephenson author page Primary source for Deepgram origin story and founder bio
Patent/press source Deepgram Series C press release and patent indexes Evidence of recent granted patents and strategic IP emphasis
Independent academic paper "How Speech Models Miss What Matters Most" Strong cautionary evidence that real-world high-stakes speech failures remain common

Two abstract geometric halves of a circuit-trace panel almost meeting, with a dashed gap between them where a waveform tries to cross

What could go wrong, and what stays private

The most important limitation is simple: Enhanced has no public modern model card in the reviewed sources. There is no official public disclosure of parameter count, context window, architecture diagram, training corpus composition, latency distribution, or a named benchmark suite for Enhanced specifically. Future AGI's public Deepgram calculator pages reinforce this absence by repeatedly marking pricing, context window, and benchmarks for Enhanced variants as not currently public or pending.

The second limitation is the documentation ambiguity in the gateway integration itself. Future AGI's formal provider docs do not document Deepgram as a provider, and the primary API reference still frames /v1/audio/transcriptions in Whisper and OpenAI-compatible terms. Meanwhile, the Deepgram cost-calculator pages clearly market Enhanced models as callable through Agent Command Center using model IDs such as deepgram/enhanced-general. Architects should treat the integration as credible but incompletely documented, not as contract-grade interface documentation.

There are broader ASR reliability concerns that apply regardless of the gateway question. Independent research shows that speech systems from top providers, including Deepgram, can still fail badly on short high-stakes utterances, non-native accented English, aphasia speech, and other off-benchmark conditions. That does not prove Enhanced is uniquely weak, but it does mean any "best-in-class STT" claim is dataset and workload dependent. Deepgram itself now publishes multiple pieces advising customers to evaluate against production audio rather than headline WER.

Privacy and compliance sit at the center of any speech stack. Deepgram's trust and self-hosting materials describe hardened systems, RBAC, encryption, SOC 2 controls, HIPAA-compatible offerings for enterprise customers, GDPR readiness with EU data residency, and self-hosted deployments that typically keep audio and transcripts inside customer infrastructure, apart from license validation and usage reporting. Future AGI makes similar control-plane claims around self-hosting, encrypted provider keys, virtual-key isolation, and caching that is optional rather than default. That is a strong operational story on paper. Any production deployment still needs an explicit review of what is cached, what metadata flows out to observability systems, and how long shadow and mirror results are retained.

Timeline and roadmap signals

Date Deepgram Future AGI Source
2015 Deepgram founded
May 2022 Enhanced launched; English availability announced
Oct 2022 Enhanced support expanded to German
2024 Public sources identify Future AGI as founded in 2024
Feb 2025 Nova-3 launched
Oct 2025 Future AGI public roundup highlights open-source stack progress
Jan 2026 Deepgram raises $130M Series C at $1.3B valuation; expands patent emphasis; acquires OfOne
Mar to May 2026 Deepgram publishes newer benchmark and voice-AI comparison materials; Enhanced still documented but not central Future AGI publishes multi-model routing comparisons, gateway docs, and Deepgram model catalog pages carrying Agent Command Center routing claims

Deepgram's roadmap signal is indirect but readable: current marketing and changelog emphasis favors Nova-3, Flux, multilingual expansion, medical specialization, and broader voice-agent stacks over any further public narrative around Enhanced. Enhanced is supported but is no longer the flagship innovation story.

Future AGI's roadmap signal is also indirect. The open-source gateway README says official Kubernetes manifests and Helm charts were "coming soon," and the surrounding 2026 blog and docs ecosystem shows ongoing investment in open-source tracing and evals, gateway breadth, and closed-loop optimization where evaluation outcomes inform routing. That points toward a platform strategy rather than a thin proxy strategy.

What remains unanswered

Several details remain unspecified in the public sources reviewed:

  • Enhanced's exact model architecture beyond "next generation end-to-end deep learning."
  • Enhanced's training corpus size, data mix, parameter count, and model-card details.
  • Public WER and latency benchmark tables for Enhanced versus current competitors.
  • A formal Future AGI provider doc page showing exactly how Deepgram is configured under Agent Command Center's provider registry.
  • Full public algorithmic descriptions for all 15 Agent Command Center routing strategies, especially providerlock, race/hedged, and some aspects of conditional and healthmonitor.
  • A complete official Deepgram founding roster beyond Scott Stephenson on the official pages reviewed.

None of these gaps is disqualifying. But they are the exact places where a proof-of-concept and a production deployment diverge, and where a week of hands-on verification will tell you more than any vendor page.

Sources

Additional primary URLs consulted:

  • https://developers.deepgram.com/reference/speech-to-text/listen-streaming
  • https://developers.deepgram.com/docs/pre-recorded-audio
  • https://developers.deepgram.com/docs/supported-audio-formats
  • https://developers.deepgram.com/docs/self-hosted-introduction
  • https://deepgram.com/pricing
  • https://deepgram.com/learn/coval-validates-flux-no-tradeoff-between-latency-and-interruption
  • https://docs.futureagi.com/docs/command-center/
  • https://docs.futureagi.com/docs/command-center/features/cost-tracking/
  • https://futureagi.com/llm-cost-calculator/deepgram/
  • https://futureagi.com/llm-cost-calculator/deepgram/enhanced/
  • https://futureagi.com/llm-cost-calculator/deepgram/enhanced-phonecall/
  • https://futureagi.com/llm-cost-calculator/deepgram/enhanced-finance/
  • https://futureagi.com/llm-cost-calculator/deepgram/enhanced-meeting/
  • https://futureagi.com/blog/best-ai-gateways-model-routing/
  • https://github.com/future-agi/agent-command-center-sdk
  • https://github.com/future-agi/traceai
  • https://arxiv.org/abs/2602.12249
  • https://patents.google.com/patent/US10540959B1/en
  • https://patents.google.com/patent/US12334075B2
  • https://www.reuters.com/technology/voice-ai-startup-deepgram-raises-130-million-13-billion-valuation-2026-01-13/
◆ The platform

Put these benchmarks to work

The same evaluations behind these dispatches drive OpenTranscription — one API that routes every job to the right speech model for your audio, language, and budget.

Try OpenTranscription →See the live leaderboard
© 2026 OpenTranscription · Signal is our journal.Set in system grotesque, serif & mono