Bland
Bland owns its whole voice stack (speech-to-text, inference, text-to-speech) and has built the testing rigour to match: evals with LLM judges, node-level regression tests, scenario suites with production gates. That engineering depth is the reason to pick it, and the reason a non-technical team should not.
PRICING
| Start | Free · $0.14/min talk time, $0.05/min transfer | 100 calls/day and /hour, 10 concurrent calls, 1 voice, 10 knowledge bases, 99.9% uptime SLA. No free minutes. |
| Build | $299/month · $0.12/min talk time, $0.04/min transfer | 2,000 calls/day, 1,000/hour, 50 concurrent calls, 5 voices, 50 knowledge bases |
| Scale | $499/month · $0.11/min talk time, $0.03/min transfer | 5,000 calls/day, 1,000/hour, 100 concurrent calls, 15 voices, 100 knowledge bases |
| Enterprise | Custom, contracted to volume | Unlimited caps, dedicated orchestration server, on-prem or VPC deployment, custom voice actor, warm and live transfers, guardrails, citations, outcomes, alarms, BAA, SSO, JWT signatures, data residency, forward-deployed engineers |
The headline $0.09/min figure quoted in older reviews is gone: per-minute rates rose across all tiers in December 2025, so the free plan now bills $0.14/min. Costs stack in several places at once — subscription, connected minutes, transfer minutes, a $0.015 minimum charged even on failed outbound attempts under 10 seconds, SMS, voice cloning, and eval runs which bill token usage per judged call. Bring-your-own-Twilio removes transfer fees. The features most SMB buyers assume are standard (guardrails, citations, warm transfers, SSO, BAA, data residency) are all Enterprise-only.
checked 2026-08-06 · vendor pricing page
Element scores
Strengths
Two things separate Bland from the voice-agent field. First, it owns the stack: proprietary speech-to-text, inference and text-to-speech on its own GPU infrastructure, so it lists no AI sub-processors in the call path, can pin every leg of a call (telephony, inference, recording, audit) to one region, and ships enterprise customers isolated Kubernetes infrastructure with version-pinned releases and canary deployments. One DPA instead of four. Second, and more unusual, it treats voice agents as software that needs tests: Evals grade real calls with weighted LLM judges and pass thresholds, Agent Testing runs simulated callers against ten assertion types with production gates that block a bad version from promotion, Standards catch node-level prompt regressions across ten permutations, and the Testbed replays a specific broken call against an edited prompt without redialling anyone. Compliance is deep for a three-year-old company: SOC 2 Type II, HIPAA with BAA, PCI DSS v4.0, and named regulated customers including CNO Financial Group and Kin Insurance.
Honest dings
Latency is the honest technical weak point: independent 2026 measurements put Bland at roughly 700-900ms average response versus 550-800ms for Vapi and Retell, and reviewers report the lag compounding over long calls, which is awkward for a company whose pitch is long complex conversations. The public review base is thin and mixed — around 3.3/5 on G2 across very few reviews — with low-rated entries citing hallucinated dollar amounts, conversational loops and refusal to transfer to a human, and multiple independent reviews saying a dedicated engineer is needed to monitor production agents after launch. Pricing stacks in at least five places and the December 2025 increase invalidated the $0.09/min figure still repeated across the web. Most of what a regulated SMB actually needs (guardrails, citations, warm transfers, BAA, SSO, data residency) is locked behind Enterprise, so the self-serve tiers are a prototyping surface, not a compliant deployment. And for a vendor that owns its weights, there is no model card or interpretability surface at all.
Sources
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- prnewswire.com/news-releases/bland-surpasses-100m-f… — Official press release, 16/06/2026: $50M Series C led by Dell Technologies Capital, $100M+ total raised, in-house models, 3.5M conversations/week, 250+ enterprise customers, 175M+ calls in the past year, named customers Samsara, Kin Insurance, CNO Financial Group, CEO Isaiah Granet quotes (accessed 2026-08-06)
- bland.ai/pricing — Official pricing page: Start/Build/Scale/Enterprise talk-time and transfer rates, daily and hourly caps, concurrency, voice and knowledge-base limits, and the full Enterprise-only feature list (guardrails, citations, warm transfers, BAA, SSO, data residency, FDE support, 28-day deployment framework) (accessed 2026-08-06)
- docs.bland.ai/tutorials/evals — Official docs: eval agents as configurable LLM judges, text vs audio modality, weights and target levels, runs up to 5,000 calls and 50 agents, workbench setups with versioning, shipped templates (Hallucination Detection, Resolution, Bland Tone, Audio Quality), cost estimation, automatic post-call evals (accessed 2026-08-06)
- docs.bland.ai/tutorials/scenarios — Official docs: Agent Testing with tester personas, ten assertion types including LLM_JUDGE and NODE_REACHED and TRAVERSAL_MATCH, production gates blocking promotion, simulation sets for flakiness detection, node failure heatmap, weakest-link and Sankey analytics, Norm auto-fix loop (accessed 2026-08-06)
- docs.bland.ai/tutorials/standards — Official docs: node-level regression framework, ten runs per standard, nine generated permutations for loop-condition and variable-extraction standards, simulation prompts for dialogue standards, pass thresholds (accessed 2026-08-06)
- docs.bland.ai/tutorials/call-logs — Official docs: transcripts, recordings, per-turn decision data with loop-condition likelihood scores, variable extraction with originating node, webhook payload and resend, tool logs, guardrail triggers, per-call cost, quality metrics panel, review workflow (accessed 2026-08-06)
- docs.bland.ai/tutorials/guard-rails — Official docs: must-trigger rails with 30-second default window for AI, self-introduction and recording disclosure (TCPA), must-not-trigger content rails, actions (end call, transfer to human, move to node), Enterprise limit of five custom rails with backtesting (accessed 2026-08-06)
- docs.bland.ai/enterprise-features/infrastructure-an… — Official docs: dedicated Kubernetes infrastructure per enterprise customer, named runtime releases with version pinning, zero-downtime rolling deployment, canary deployments with percentage and rule-based traffic routing (accessed 2026-08-06)
- app.bland.com/trust-security — Official trust page: SOC 2 Type II, HIPAA, PCI DSS v4.0, no AI sub-processors in the call path, per-call regional pinning with EU data on EU GPUs, Bland cloud / VPC / on-premise deployment options, quarterly third-party penetration tests, claim that guardrails are fine-tuned into model weights (accessed 2026-08-06)
- docs.bland.ai/sdks/cli — Official docs: CLI covering calls, pathways as YAML with push/pull/diff/validate/watch, personas with gaps and promote, tools with test invocation, knowledge bases with scraping, batches, guard rails, alarms, secrets, releases, widgets, evals; local stdio MCP server plus the hosted remote server at api.bland.ai/v1/mcp (accessed 2026-08-06)
- docs.bland.ai/tutorials/norm — Official docs: Norm builds, edits, tests and debugs agents from a prompt, forks a draft so production keeps running, replays real calls through Test Bed, requires an explicit promotion step to go live, supports custom team skills (accessed 2026-08-06)
- docs.bland.ai/tutorials/personas — Official docs: personas as unified agents routing across pathways by context, knowledge-base connection, draft/production versioning with diffs, attached evals and outcomes, Enterprise-only data extraction with citation schemas (accessed 2026-08-06)
- bland.ai/legal/model-governance — Official governance brief: customer-exclusive fine-tuning, LoRA hyperparameter disclosure, adapter layers, quantisation-aware training, exclusive voice-actor licensing, data-sovereignty and no-training commitments, all framed as an enterprise security-review conversation (accessed 2026-08-06)
- docs.bland.ai/enterprise-features/SSO — Official docs: OIDC and SAML 2.0 SSO with Okta, Azure AD, Google Workspace and Auth0; just-in-time provisioning with SCIM explicitly still on the roadmap (accessed 2026-08-06)
- thoughtly.com/blog/retell-vs-bland-comparison — Independent competitor comparison, 2026: Bland at roughly 3.3/5 on G2 with limited reviews versus Retell's 4.8/5 across 2,200+, low-rated reviews citing hallucinated information, conversational loops and refusal to transfer, and the recurring finding that a dedicated engineer is needed to monitor production agents (accessed 2026-08-06)
- coval.ai/blog/bland-ai-review-2026-features-pricing… — Independent 2026 review: measured latency of 700-900ms versus Vapi 550-800ms and Retell 600-800ms, outbound optimisation, pathway structure trading flexibility for consistency, campaign-level rather than conversation-level monitoring (accessed 2026-08-06)
- autocalls.ai/article/bland-ai-review — Independent 2026 pricing analysis documenting the December 2025 per-minute increase (Start from $0.09 to $0.14, a 55% rise), transfer fees, the $0.015 minimum on failed outbound attempts, and voicemail billing (accessed 2026-08-06)
- app.dealroom.co/companies/bland_ai — Independent funding database corroborating the round history ($125k seed 06/2023, $16M Series A 08/2024, $40M Series B 02/2025, $50M Series C 06/2026), founders, and San Francisco base (accessed 2026-08-06)
- ycombinator.com/companies/bland-ai — Official YC company profile and launch post: enterprise platform for AI phone calls, call-centre market framing, plus the Fortune headline that the Series C came after rejection by 180 investors (accessed 2026-08-06)