Command A+

v20260520

Cohere

Modelenterpriseragcitationsopen-source
87
Strong
About This Model

Cohere's Apache 2.0 open-weight 218B sparse MoE (25B active) unifying Command A, A Reasoning, A Vision, and A Translate. Runs on 2xH100 or a single B200, supports 48 languages, and ships native citations with grounding spans for verifiable RAG.

Last Evaluated: June 10, 2026
Official Website

Trust Vector Analysis

Dimension Breakdown

🚀Performance & Reliability
+

Released 2026-05-20, unifying Command A + A Reasoning + A Vision + A Translate into one model. Efficient serving (25B active, 2xH100 or 1xB200). Benchmarks are vendor-reported and independent verification is still emerging.

task accuracy code

Vendor-reported coding benchmarks; independent replication still emerging three weeks post-release

Evidence
Command A+ launch announcementSolid coding performance; competitive with open peers but not the headline focus
mediumVerified: 2026-06-10
task accuracy reasoning

Vendor-reported reasoning benchmarks pending broad independent verification

Evidence
Command A+ launch announcementUnifies Command A Reasoning capabilities; multimodal reasoning over text and images
mediumVerified: 2026-06-10
task accuracy general

Review of vendor benchmarks emphasizing enterprise RAG, translation, and multilingual tasks

Evidence
Command A+ launch announcementStrong enterprise-task performance across 48 languages; unifies four specialist Command A variants
mediumVerified: 2026-06-10
output consistency

Review of grounded-generation behavior and enterprise consistency claims

Evidence
Cohere DocumentationEnterprise focus on reproducible, grounded outputs; structured citation format is deterministic
mediumVerified: 2026-06-10
latency p50

Median latency for API requests with standard prompt sizes

Evidence
Community benchmarkingTypical response ~2.0s; 25B active parameters keep serving fast on 2xH100
mediumVerified: 2026-06-10
latency p95

95th percentile response time across diverse workloads

Evidence
Community benchmarkingp95 ~4.5s across diverse workloads
mediumVerified: 2026-06-10
context window

Provider documentation; A+ figure carried from Command A line pending explicit spec sheet

Evidence
Cohere model documentation256K context consistent with prior Command A; A+ spec inferred pending explicit confirmation
mediumVerified: 2026-06-10
uptime

Historical uptime data from official status page

Evidence
Cohere Status PageStrong historical availability; private/VPC deployment removes dependence on shared API
highVerified: 2026-06-10
🛡️Security
+

Strongest security posture among current open-weight releases thanks to Cohere's enterprise platform: SOC 2, VPC/private deployment, and a grounded-generation design that narrows injection surface in RAG.

prompt injection resistance

Review of vendor security documentation and enterprise deployment guidance against OWASP LLM01 patterns

Evidence
Cohere security documentationEnterprise hardening program; grounded-RAG design constrains injection surface in retrieval workflows
mediumVerified: 2026-06-10
jailbreak resistance

Testing against adversarial prompt datasets; deployer-dependent for self-hosted use

Evidence
Cohere responsible AI documentationEnterprise-grade safety tuning; open weights mean self-hosted derivatives can strip guardrails
mediumVerified: 2026-06-10
data leakage prevention

Analysis of deployment isolation options and data handling practices

Evidence
Cohere enterprise deployment optionsPrivate deployment and VPC options keep data fully inside customer infrastructure
highVerified: 2026-06-10
output safety

Safety testing across harmful content categories

Evidence
Cohere responsible AI documentationSafety filtering tuned for enterprise content categories
mediumVerified: 2026-06-10
api security

Review of API security features, certifications, and enterprise controls

Evidence
Cohere security pageSOC 2 audited platform, API key auth, HTTPS only, SSO/SCIM for enterprise, VPC endpoints
highVerified: 2026-06-10
🔒Privacy & Compliance
+

Best-in-class privacy posture for an open-weight model: Western jurisdiction, SOC 2, and a full spectrum from SaaS to on-premises. The rare combination of open weights plus enterprise compliance is its core differentiator.

data residency

Review of deployment documentation and residency options

Evidence
Cohere deployment optionsSaaS, cloud-marketplace, VPC, and on-premises deployment; Apache 2.0 weights allow any-jurisdiction self-hosting
highVerified: 2026-06-10
training data optout

Analysis of privacy policy and enterprise data usage terms

Evidence
Cohere privacy policyEnterprise data not used for training by default under enterprise terms
highVerified: 2026-06-10
data retention

Review of terms of service and deployment-dependent retention

Evidence
Cohere enterprise documentationPrivate deployments process data entirely within customer environment
highVerified: 2026-06-10
pii handling

Review of data protection capabilities and customer responsibilities

Evidence
Cohere security documentationEnterprise data-handling guidance; customer retains redaction responsibility, but isolation options reduce exposure
mediumVerified: 2026-06-10
compliance certifications

Verification of compliance certifications and audit reports

Evidence
Cohere security and trust pageSOC 2 Type II, GDPR alignment, enterprise agreements; long-standing enterprise compliance posture
highVerified: 2026-06-10
zero data retention

Review of self-hosting and private deployment options enabling zero retention

Evidence
Apache 2.0 open weightsSelf-hosting on 2xH100 or 1xB200 gives complete data control; VPC option for managed zero-retention
highVerified: 2026-06-10
👁️Trust & Transparency
+

Native citations and grounding spans are a genuine trust differentiator: outputs in RAG mode are verifiable against source passages by construction, not post-hoc.

explainability

Evaluation of citation fidelity and grounding-span behavior in retrieval workflows

Evidence
Command A+ launch announcementNative citations with grounding spans tie each claim to source passages — verifiable RAG by construction
highVerified: 2026-06-10
hallucination rate

Testing on grounded QA workloads; ungrounded closed-book use performs closer to peers

Evidence
Cohere grounded generation documentationGrounded mode constrains generation to retrieved evidence, materially reducing unsupported claims
mediumVerified: 2026-06-10
bias fairness

Review of published bias evaluations and responsibility framework

Evidence
Cohere responsible AI documentationPublished responsibility framework and multilingual fairness work across 48 languages
mediumVerified: 2026-06-10
uncertainty quantification

Assessment of confidence expression and grounding-span signaling

Evidence
Grounding span behaviorAbsence of grounding spans signals unsupported content, giving a usable uncertainty proxy in RAG
mediumVerified: 2026-06-10
model card quality

Review of documentation completeness and clarity

Evidence
Command A+ release materialsClear documentation of architecture (218B sparse MoE / 25B active), hardware targets, languages, and unified capabilities
highVerified: 2026-06-10
training data transparency

Review of public disclosures about training data

Evidence
Cohere public materialsGeneral training approach described; detailed data sources not disclosed (industry standard)
mediumVerified: 2026-06-10
guardrails

Analysis of built-in safety mechanisms and configurability

Evidence
Cohere responsible AI documentationEnterprise safety tuning with configurable safety modes
mediumVerified: 2026-06-10
⚙️Operational Excellence
+

Mature enterprise platform behind a permissively licensed model. Apache 2.0 weights plus SOC 2 SaaS/VPC options give an unusually wide deployment spectrum.

api design quality

Review of API design, consistency, and feature completeness

Evidence
Cohere API DocumentationMature Chat API with first-class RAG primitives (documents, citations), tool use, and streaming
highVerified: 2026-06-10
sdk quality

Review of SDK quality, documentation, and maintenance

Evidence
Cohere SDKsOfficial Python, TypeScript, Go, and Java SDKs, actively maintained
highVerified: 2026-06-10
versioning policy

Review of versioning policy and historical practices

Evidence
Cohere model documentationStable model naming with deprecation notices; A+ consolidates four variants into one maintained line
mediumVerified: 2026-06-10
monitoring observability

Review of available monitoring tools and metrics

Evidence
Cohere dashboardUsage dashboards and logs; enterprise deployments integrate with customer observability stacks
mediumVerified: 2026-06-10
support quality

Assessment of documentation, community, and support responsiveness

Evidence
Cohere supportDedicated enterprise support, solutions engineering, and comprehensive documentation
highVerified: 2026-06-10
ecosystem maturity

Analysis of third-party integrations, cloud availability, and tooling

Evidence
Cloud availabilityAvailable via Cohere API, major cloud marketplaces, and Hugging Face weights for self-hosting
highVerified: 2026-06-10
license terms

Review of licensing terms and restrictions

Evidence
Command A+ launch announcementApache 2.0 open weights — fully permissive commercial use, a notable shift from Cohere's earlier non-commercial open releases
highVerified: 2026-06-10
Strengths
  • +Native citations with grounding spans — verifiable RAG outputs by construction
  • +Apache 2.0 open weights from a Western enterprise vendor with SOC 2 and VPC/on-prem options
  • +Efficient serving: 218B sparse MoE with 25B active, runs on 2xH100 or a single B200
  • +Unifies Command A, A Reasoning, A Vision, and A Translate into one multimodal model
  • +48-language support with strong enterprise translation quality
  • +Full deployment spectrum: SaaS, cloud marketplace, VPC, on-premises, self-hosted
Limitations
  • !Command A+ pricing not clearly published as of 2026-06-10 (prior Command A was $2.50/$10 per 1M)
  • !Benchmarks vendor-reported; independent verification still emerging three weeks post-release
  • !Behind agentic-coding specialists on SWE-bench-style tasks
  • !Training data sources not disclosed in detail
  • !Smaller open-source community than Chinese open-weight ecosystems
Metadata
pricing
input: Not yet published (prior Command A: $2.50 per 1M tokens)
output: Not yet published (prior Command A: $10.00 per 1M tokens)
notes: Command A+ API pricing not clearly published as of 2026-06-10; figures from the prior Command A generation shown for reference. Low confidence — verify with Cohere before procurement.
last verified: 2026-06-10
context window: 256000
languages
0: English
1: French
2: German
3: Spanish
4: Italian
5: Portuguese
6: Japanese
7: Korean
8: Chinese
9: Arabic
10: Hindi
modalities
0: text
1: image (input)
api endpoint: https://api.cohere.com/v2/chat
open source: true
license: Apache 2.0
architecture: Sparse Mixture-of-Experts: 218B total / 25B active parameters; multimodal reasoning; unified Command A + Reasoning + Vision + Translate
parameters: 218B total / 25B active
release date: 2026-05-20
supported language count: 48

Use Case Ratings

code generation

Capable but not the focus; agentic-coding specialists (Kimi K2.6, GLM-5) score higher.

customer support

Excellent fit: 48 languages, grounded answers with citations, fast 25B-active serving, enterprise deployment options.

content creation

Strong multilingual content with citation support for sourced writing.

data analysis

Solid analysis with multimodal input and verifiable grounding over enterprise documents.

research assistant

Native citations with grounding spans make research outputs auditable; 256K context handles large corpora.

legal compliance

Western jurisdiction, SOC 2, VPC/on-prem options, and verifiable citations suit legal document workflows.

healthcare

Private/self-hosted deployment supports strict data control; verify HIPAA terms for managed offerings.

financial analysis

Grounded generation over filings and reports with citation trails fits compliance-sensitive finance work.

education

Strong multilingual tutoring; citations encourage source-checking habits.

creative writing

Serviceable but enterprise-tuned; creative specialists perform better.